Methodological and Conceptual Framework for 5D Multi-Table Analysis: A Unified Approach for Complex Data Reuse

# Methodological and Conceptual Framework for 5D Multi-Table Analysis: A Unified Approach for Complex Data Reuse
Source: [https://arxiv.org/html/2608.26149](https://arxiv.org/html/2608.26149)
Hugo Kazzi Lille Centrale Institute hugo\.kazzi@master\.centralelille\.comAurélien Loison Lille Centrale Institute aurelien\.loison\.etu@univ\-lille\.frPr\. Slim Hammadi Lille Centrale Institute, CRISTAL UMR CNRS 9189 slim\.hammadi@centralelille\.comPr\. Emmanuel Chazard Department of Public Health, EA 2694, ULR 2694\-METRICS, Lille University emmanuel\.chazard@chu\-lille\.fr

###### Abstract

Multi\-table analysis remains a major challenge in machine learning applied to healthcare and complex information systems\. Relational data simultaneously present five dimensions of complexity: massive volume, multiplicity of variables, high cardinality of categorical variables, complex inter\-table relationships, and repeated temporal measurements\. We propose an integrated architecture, theRelational Hypergraph Transformer\(RHT\), that addresses these dimensions through a unified representation combining relational hypergraphs, pentadimensional embeddings \(PentE\), and a sparse relational attention mechanism whose cost scales asO​\(n⋅k\)O\(n\\cdot k\)in the average relational degreekkrather thanO​\(n2\)O\(n^\{2\}\)in the number of entitiesnn\. We formalise the architecture, prove the complexity bound of its core attention operator, and release an open\-source reference implementation\.

We evaluate the approach empirically on the openSyntheasynthetic electronic\-health\-record dataset \(SNOMED CT condition codes\), on the task of multi\-label condition code prediction per encounter — a setting that exercises high categorical cardinality with a long\-tailed code distribution\. We compare against tabular \(XGBoost\), relational \(GraphSAGE\), and temporal\-graph baselines, and report rare\-category recall \(RCR@kk\), macro\-F1, and embedding semantic coherence, together with an ablation isolating the contribution of each module\. RHT improves semantic coherence \(1\.52±\\pm0\.03\) relative to non\-hierarchical relational baselines, while the strongest rare\-code recall in this setting is achieved by XGBoost\. We position MIMIC\-IV as a planned clinical validation contingent on PhysioNet credentialing \(Phase 1 of our validation plan\)\. Code and experiments are available at the repository listed in Appendix B\.

Scope\.This paper contributes \(i\) a formal architecture with a proven complexity bound, \(ii\) an open implementation, and \(iii\) an empirical validation of the high\-cardinality components \(Modules 1 and 3\) on synthetic EHR data\. The temporal and relational\-discovery modules \(Modules 2 and 4\) are specified and implemented but their large\-scale clinical validation is deferred to ongoing work\. We are explicit throughout about which results are measured and which are targets for future validation\.

Keywords:Multi\-table learning, relational learning, graph neural networks, high cardinality, electronic health records, digital health

###### Contents

1. [1Introduction](https://arxiv.org/html/2608.26149#S1)1. [1\.1Context and Motivation](https://arxiv.org/html/2608.26149#S1.SS1) 2. [1\.2Research Problem](https://arxiv.org/html/2608.26149#S1.SS2) 3. [1\.3The Five Dimensions of Complexity](https://arxiv.org/html/2608.26149#S1.SS3) 4. [1\.4Contributions of This Work](https://arxiv.org/html/2608.26149#S1.SS4) 5. [1\.5Paper Structure](https://arxiv.org/html/2608.26149#S1.SS5)
2. [2State of the Art and Comparative Analysis](https://arxiv.org/html/2608.26149#S2)1. [2\.1Taxonomy of Existing Approaches](https://arxiv.org/html/2608.26149#S2.SS1)1. [2\.1\.1Classical Approaches](https://arxiv.org/html/2608.26149#S2.SS1.SSS1) 2. [2\.1\.2Traditional ML Approaches](https://arxiv.org/html/2608.26149#S2.SS1.SSS2) 3. [2\.1\.3Modern Deep Learning Approaches](https://arxiv.org/html/2608.26149#S2.SS1.SSS3) 2. [2\.2Identified Gaps in Literature](https://arxiv.org/html/2608.26149#S2.SS2)
3. [3Prospective Vision and Conceptual Framework](https://arxiv.org/html/2608.26149#S3)1. [3\.1Strategic Vision](https://arxiv.org/html/2608.26149#S3.SS1) 2. [3\.2Overview of the Unified Architecture](https://arxiv.org/html/2608.26149#S3.SS2) 3. [3\.3Functional Layers of the Meta\-Model \(MTAM\)](https://arxiv.org/html/2608.26149#S3.SS3)
4. [4Relational Hypergraph Transformer Architecture](https://arxiv.org/html/2608.26149#S4)1. [4\.1Motivation and Design Principles](https://arxiv.org/html/2608.26149#S4.SS1) 2. [4\.2Computational Modules](https://arxiv.org/html/2608.26149#S4.SS2)1. [4\.2\.1Module 1: Hypergraph Construction](https://arxiv.org/html/2608.26149#S4.SS2.SSS1) 2. [4\.2\.2Module 2: Multi\-Scale Temporal Embeddings](https://arxiv.org/html/2608.26149#S4.SS2.SSS2) 3. [4\.2\.3Module 3: High\-Cardinality Attention](https://arxiv.org/html/2608.26149#S4.SS2.SSS3) 4. [4\.2\.4Module 4: Differentiable Relational Discovery](https://arxiv.org/html/2608.26149#S4.SS2.SSS4)
5. [5Eight\-Step Methodology](https://arxiv.org/html/2608.26149#S5)1. [5\.1Step 1: Predictive Meta\-Profiling](https://arxiv.org/html/2608.26149#S5.SS1) 2. [5\.2Step 2: Automated Hypergraph Construction](https://arxiv.org/html/2608.26149#S5.SS2) 3. [5\.3Step 3: Unified 5D Embedding \(PentE\)](https://arxiv.org/html/2608.26149#S5.SS3) 4. [5\.4Step 4: Relational Contrastive Learning](https://arxiv.org/html/2608.26149#S5.SS4) 5. [5\.5Step 5: Dynamic Graph Rewiring](https://arxiv.org/html/2608.26149#S5.SS5) 6. [5\.6Step 6: Relational Causal Inference](https://arxiv.org/html/2608.26149#S5.SS6) 7. [5\.7Step 7: Federated Multi\-Table Learning](https://arxiv.org/html/2608.26149#S5.SS7) 8. [5\.8Step 8: Operationalization and Monitoring](https://arxiv.org/html/2608.26149#S5.SS8)
6. [6Key Technical Innovations](https://arxiv.org/html/2608.26149#S6)1. [6\.1Sparse Relational Attention](https://arxiv.org/html/2608.26149#S6.SS1) 2. [6\.2Hierarchical High\-Cardinality Encoding](https://arxiv.org/html/2608.26149#S6.SS2) 3. [6\.3Temporal\-Relational Message Passing](https://arxiv.org/html/2608.26149#S6.SS3) 4. [6\.4Causal Relational Discovery](https://arxiv.org/html/2608.26149#S6.SS4)
7. [7Comparative Benchmark and Evaluation](https://arxiv.org/html/2608.26149#S7)1. [7\.1Reference Datasets](https://arxiv.org/html/2608.26149#S7.SS1)1. [7\.1\.1MIMIC\-IV \(Medical Information Mart for Intensive Care\)](https://arxiv.org/html/2608.26149#S7.SS1.SSS1) 2. [7\.1\.2Amazon Multi\-Table Dataset](https://arxiv.org/html/2608.26149#S7.SS1.SSS2) 3. [7\.1\.3Financial Transactions Dataset](https://arxiv.org/html/2608.26149#S7.SS1.SSS3) 2. [7\.2Multidimensional Evaluation Metrics](https://arxiv.org/html/2608.26149#S7.SS2)1. [7\.2\.1Dimension 1: Volume \(Scalability\)](https://arxiv.org/html/2608.26149#S7.SS2.SSS1) 2. [7\.2\.2Dimension 2: Many Variables](https://arxiv.org/html/2608.26149#S7.SS2.SSS2) 3. [7\.2\.3Dimension 3: High Cardinality](https://arxiv.org/html/2608.26149#S7.SS2.SSS3) 4. [7\.2\.4Dimension 4: Multiple Tables](https://arxiv.org/html/2608.26149#S7.SS2.SSS4) 5. [7\.2\.5Dimension 5: Repeated Measurements](https://arxiv.org/html/2608.26149#S7.SS2.SSS5) 6. [7\.2\.6Anomaly and Extreme Value Detection](https://arxiv.org/html/2608.26149#S7.SS2.SSS6) 7. [7\.2\.7Holistic Metrics](https://arxiv.org/html/2608.26149#S7.SS2.SSS7) 3. [7\.3Experimental Setup](https://arxiv.org/html/2608.26149#S7.SS3) 4. [7\.4Results](https://arxiv.org/html/2608.26149#S7.SS4) 5. [7\.5Ablation](https://arxiv.org/html/2608.26149#S7.SS5) 6. [7\.6Phase 1 Pilot on Existing Benchmark Data \(MIMIC\-IV Demo\)](https://arxiv.org/html/2608.26149#S7.SS6) 7. [7\.7Asymptotic Scaling \(Analytical\)](https://arxiv.org/html/2608.26149#S7.SS7) 8. [7\.8Progressive Validation Plan](https://arxiv.org/html/2608.26149#S7.SS8) 9. [7\.9Current Status and Remaining Phase](https://arxiv.org/html/2608.26149#S7.SS9) 10. [7\.10Phase 2 Progress Achieved Without PhysioNet Access](https://arxiv.org/html/2608.26149#S7.SS10)
8. [8Implications and Perspectives](https://arxiv.org/html/2608.26149#S8)1. [8\.1Toward a Formal Theory of Learning on Temporal Hypergraphs](https://arxiv.org/html/2608.26149#S8.SS1) 2. [8\.2Establishing Relational Data Science as an Interdisciplinary Research Program](https://arxiv.org/html/2608.26149#S8.SS2) 3. [8\.3Industrial Implications and Operational Transformation](https://arxiv.org/html/2608.26149#S8.SS3) 4. [8\.4Evolution of Analytical Tools and Platforms](https://arxiv.org/html/2608.26149#S8.SS4)
9. [9Conclusion](https://arxiv.org/html/2608.26149#S9)1. [9\.1Summary of Contributions](https://arxiv.org/html/2608.26149#S9.SS1) 2. [9\.2Decisive Advantages of the Proposed Approach](https://arxiv.org/html/2608.26149#S9.SS2) 3. [9\.3Limitations and Directions for Future Research](https://arxiv.org/html/2608.26149#S9.SS3) 4. [9\.4Research Roadmap](https://arxiv.org/html/2608.26149#S9.SS4) 5. [9\.5General Conclusion](https://arxiv.org/html/2608.26149#S9.SS5)
10. [References](https://arxiv.org/html/2608.26149#bib)
11. [ADetailed Pseudo\-code](https://arxiv.org/html/2608.26149#A1)1. [A\.1Hypergraph Construction](https://arxiv.org/html/2608.26149#A1.SS1) 2. [A\.2PentE Embeddings](https://arxiv.org/html/2608.26149#A1.SS2)
12. [BDatasets and Code](https://arxiv.org/html/2608.26149#A2)

## 1Introduction

### 1\.1Context and Motivation

The explosion of digital data in healthcare, finance, e\-commerce, and Internet of Things \(IoT\) sectors has created an urgent need for analytical methods capable of efficiently processing complex and heterogeneous data structures\. Unlike traditional tabular data, these environments generate data distributed across multiple interconnected tables, with intrinsic relationships carrying critical meaning and information\.

In the medical domain, for example, electronic health records \(EHR\) typically comprise dozens of linked tables: patient demographics, diagnostic histories, medication prescriptions, laboratory results, and repeated physiological measurements\. Extracting actionable knowledge from these data requires a holistic understanding of their relational structure, temporal evolution, and semantic richness\.

### 1\.2Research Problem

Traditional data analysis approaches rely on transforming relational structures into unified wide tables via SQL join operations\. This strategy presents several major limitations\. First, joining multiple tables generates an explosion in the number of columns and rows, making processing computationally prohibitive—a problem that compounds with each additional table in the schema\. Second, the structure of relationships between entities, which carries essential information about how data are generated and interconnected, is flattened and lost in the merge process; this is particularly detrimental when relational semantics encode domain knowledge \(e\.g\., the clinical pathway linking a patient’s admission to diagnoses, prescriptions, and outcomes\)\. Third, repeated measurements and irregular time series are poorly represented in classic tabular structures, as flattening temporal sequences into columns either discards ordering information or creates extremely sparse representations\. Fourth, categorical variables with thousands of possible values \(diagnostic codes, product identifiers\) pose major challenges to traditional machine learning algorithms, which typically assume moderate cardinality for one\-hot encoding or similar strategies\.

This problem intensifies when data simultaneously present five dimensions of complexity, as illustrated in the context of medicalbig data\.

### 1\.3The Five Dimensions of Complexity

Our conceptual framework is structured around five critical dimensions that characterize the complexity of modern multi\-table data\.Dimension 1: Massive Volumerefers to the fact that the number of records and tables can reach orders of magnitude that make classical approaches impractical, requiring methods that scale sublinearly with data size\.Dimension 2: Multiplicity of Variablescaptures the challenge that each table can contain dozens or even hundreds of columns, requiring intelligent dimensional reduction strategies that preserve inter\-variable dependencies\.Dimension 3: High Categorical Cardinalityaddresses the challenge posed by categorical variables with thousands of distinct values \(e\.g\., ICD\-10 codes in medicine\), often with heavily imbalanced distributions where rare categories carry critical diagnostic significance\.Dimension 4: Multiple Tables and Relationshipsreflects the relational architecture comprising numerous interconnected tables via complex relationships \(one\-to\-many, many\-to\-many, hierarchical\) whose structure encodes essential domain knowledge\.Dimension 5: Repeated Temporal Measurementsaccounts for variables measured repeatedly at irregular intervals, introducing an essential temporal dimension that interacts with all other dimensions\.

### 1\.4Contributions of This Work

The main contributions of this paper are fivefold\. First, we articulate aprospective visionandunified conceptual frameworkfor multi\-table analysis that, unlike existing approaches, explicitly integrates all five dimensions of complexity within a single coherent formalism\. Second, we propose anovel architecturecalledRelational Hypergraph Transformer\(RHT\) that combines hypergraph representations for n\-ary relationships, pentadimensional embeddings for unified latent\-space encoding, and adaptive attention mechanisms for efficient cross\-table learning—addressing the theoretical gap between graph\-based and temporal models identified in the state of the art\. Third, we introducemethodological innovationsincluding relational contrastive learning, dynamic graph rewiring, and relational causal inference, which extend multi\-table analysis beyond supervised prediction to unsupervised structure discovery and causal reasoning\. Fourth, we provide acomparative benchmarkwith a formal scoring methodology that comprehensively positions our approach against the state of the art \(Graph Neural Networks, Temporal Graph Networks, Statistical Relational Learning\) across all five dimensions\. Fifth, we present avalidation planincluding the creation of a new standardized benchmark \(MT\-5D\-Bench\) and multidimensional evaluation metrics designed to penalize approaches that neglect any single dimension of complexity\.

### 1\.5Paper Structure

The remainder of this paper is organized as follows\. Section[2](https://arxiv.org/html/2608.26149#S2)presents a detailed state of the art of existing approaches and their limitations\. Section[3](https://arxiv.org/html/2608.26149#S3)exposes our conceptual framework and prospective vision\. Section[4](https://arxiv.org/html/2608.26149#S4)details the unified architecture of the Relational Hypergraph Transformer, structured into four functional layers, each implemented by a dedicated computational module and orchestrated through an eight\-step methodology\. Section[6](https://arxiv.org/html/2608.26149#S6)presents key technical innovations that cut across multiple modules\. Section[7](https://arxiv.org/html/2608.26149#S7)proposes a comparative benchmark with the state of the art\. Section[7\.2](https://arxiv.org/html/2608.26149#S7.SS2)defines evaluation metrics, experimental methodology, and measured results, and clarifies the remaining validation phase\. Finally, Section[9](https://arxiv.org/html/2608.26149#S9)concludes and proposes future research perspectives\.

## 2State of the Art and Comparative Analysis

### 2\.1Taxonomy of Existing Approaches

Multi\-table analysis has been addressed from different angles in the scientific literature\. We propose a structured taxonomy in three main categories: classical approaches, traditional machine learning methods, and modern deep learning approaches\.

#### 2\.1\.1Classical Approaches

##### Table Joining \(Table Fusion\)

The most widespread approach in practice consists of merging relational tables via SQL join operations, following star schema or snowflake schema architectures proposed by Kimball\[[1](https://arxiv.org/html/2608.26149#bib.bib1)\]\. This strategy generates a single wide table that can then be processed by standard machine learning algorithms\.

Limitations:This approach suffers from several critical problems\. First, the dimensional explosion resulting from joining multiple tables creates extremely sparse and voluminous matrices\. Second, the semantics of relationships between entities is lost in the flattening process\. Third, this method fails against dimension 3 \(high cardinality\) and dimension 5 \(repeated measurements\) of our framework\.

##### Data Integration Frameworks

Platforms like Apache Atlas, CloverDX, and Talend offer solutions for integration and metadata management\. These tools excel in data lineage tracing and ETL pipeline orchestration\.

Analytical gap:These frameworks are primarily oriented toward integration and batch processing, without advanced analytics or integrated machine learning capabilities\.

#### 2\.1\.2Traditional ML Approaches

##### Basic Multi\-Table Learning

Singh and Gordon\[[2](https://arxiv.org/html/2608.26149#bib.bib2)\]proposedCollective Matrix Factorization, an extension of matrix factorization allowing simultaneous factorization of multiple linked matrices\. This approach exploits relationships between tables to improve factorization quality\.

Limits:While promising, this method does not efficiently handle high cardinality of categorical variables or temporality of repeated measurements\. Moreover, it assumes linear relationships between entities\.

##### Statistical Relational Learning \(SRL\)

Markov Logic Networks\(MLN\), introduced by Richardson and Domingos\[[3](https://arxiv.org/html/2608.26149#bib.bib3)\], combine first\-order logic and probabilistic graphical models to model relational dependencies\.

Strengths:Ability to explicitly represent complex relationships and perform probabilistic inference\.

Weaknesses:Not scalable for big data, high computational complexity, unsuitable for continuous temporal data\.

#### 2\.1\.3Modern Deep Learning Approaches

##### Graph Neural Networks \(GNN\)

Graph neural networks have revolutionized structured data processing\. Foundational architectures includeMessage Passing Networks\[[4](https://arxiv.org/html/2608.26149#bib.bib4)\]and GraphSAGE\[[6](https://arxiv.org/html/2608.26149#bib.bib6)\]\.

Multi\-table application:GNNs can represent tables as nodes and relationships as edges, enabling learning of representations that preserve relational structure\.

Deficit:Standard GNNs are not designed to handle temporality and high cardinality simultaneously\. They generally process static graphs with moderate\-dimensional node attributes\.

##### Temporal Graph Networks

To address the temporal limitation of GNNs, architectures like TGN\[[7](https://arxiv.org/html/2608.26149#bib.bib7)\]and TGAT\[[8](https://arxiv.org/html/2608.26149#bib.bib8)\]have been proposed\. These models explicitly integrate the temporal dimension using temporal attention mechanisms and memories\.

Advancement:Partial handling of dimensions 4 \(relationships\) and 5 \(temporality\) of our framework\.

Gap:Absence of specific optimization for high cardinality categorical variables and very numerous variables\.

##### Relational Transformers

Peters et al\.\[[9](https://arxiv.org/html/2608.26149#bib.bib9)\]adapted the Transformer architecture for relational data, enabling attention over relationships between entities\.

Capability:Flexible attention mechanism on relationships\.

Limitation:QuadraticO​\(n2\)O\(n^\{2\}\)complexity prohibitive for data with numerous variables and records\.

##### Qualitative capability comparison\.

Table[1](https://arxiv.org/html/2608.26149#S2.T1)summarises, for each family of methods, which of the five complexity dimensions it*natively supports*\(✓\), supports only*partially*\(∼\\sim\), or does*not*address \(×\\times\)\. These assignments reflect the formal capabilities of each method as described in the cited literature; they are qualitative and are not a substitute for the empirical comparison reported in Section[7\.2](https://arxiv.org/html/2608.26149#S7.SS2)\. We deliberately avoid assigning numerical capability scores, as such scores would imply a precision not warranted by a literature\-based assessment\.

Table 1:Qualitative native support of each method family across the five dimensions\. ✓ = native support;∼\\sim= partial support;×\\times= not addressed\. Based on the formal capabilities described in the cited references\.MethodD1D2D3D4D5Vol\.Var\.Card\.TablesTemp\.RHT \(this work\)✓✓✓✓∼\\simTGN\[[7](https://arxiv.org/html/2608.26149#bib.bib7)\]∼\\sim∼\\sim×\\times✓✓GraphSAGE\+LSTM\[[6](https://arxiv.org/html/2608.26149#bib.bib6)\]✓∼\\sim×\\times✓∼\\simCollective Matrix Fact\.\[[2](https://arxiv.org/html/2608.26149#bib.bib2)\]∼\\sim✓×\\times✓×\\timesMarkov Logic Networks\[[3](https://arxiv.org/html/2608.26149#bib.bib3)\]×\\times∼\\sim∼\\sim✓×\\timesSQL Wide Table \+ XGBoost∼\\sim×\\times×\\times×\\times×\\times
The temporal support of RHT is marked partial \(∼\\sim\): the multi\-scale temporal module is implemented but, in the present evaluation, the temporal dimension is exercised only through encounter timestamps on Synthea rather than dense irregular physiological time series\. Full temporal validation is part of the MIMIC\-IV phase\.

### 2\.2Identified Gaps in Literature

Our analysis reveals four major gaps in the current state of the art that collectively motivate the development of our unified framework\. The most fundamental gap is the absence of a holistic approach: no existing method simultaneously addresses the five dimensions of complexity\. Approaches typically focus on one or two dimensions—TGN on relationships and temporality, Collective Matrix Factorization on relationships and variables—leaving the remaining dimensions unaddressed and creating performance bottlenecks when applied to real\-world multi\-table data\. A second gap concerns the lack of standardized benchmarks: there is no recognized benchmark for comparative evaluation of multi\-table analysis methods, making objective comparison of approaches difficult and impeding reproducible research progress\. Third, no existing formalism coherently combines relational structure, temporality, and high cardinality attributes in a single representation space; current approaches require separate representation stages that introduce information loss at each transition\. Fourth, solutions developed in specific domains \(healthcare, finance, e\-commerce\) remain domain\-specific ad\-hoc constructions that lack generalization and transferability, precluding the emergence of general principles applicable across relational data domains\.

These gaps motivate the development of our unified conceptual and methodological framework, presented in the following sections\.

## 3Prospective Vision and Conceptual Framework

### 3\.1Strategic Vision

Our vision is to create aunified analytical ecosystemwhere complex relational data are treated as a holistic system rather than a collection of disjoint tables\. This ecosystem must enable organizations to discoverinsightsacross natural tabular boundaries through contextual and relational intelligence\.

This vision rests on three fundamental pillars:

##### Transparent Semantic Integration\.

The system’s capability to automatically understand relationships between entities without extensive human intervention, exploiting both explicit constraints \(foreign keys\) and latent semantic dependencies\.

##### Cognitive Abstraction\.

A unified representation of multi\-table data in a latent space that preserves relational semantics while enabling efficient analytical operations\.

##### Relational Inference\.

The automatic discovery of trans\-tabular patterns and knowledge by jointly exploiting relational structure, entity attributes, and their temporal evolution\.

### 3\.2Overview of the Unified Architecture

Before describing each component in detail, we present the overall structure of our solution\. The framework is organized into three levels of description that serve complementary roles\.Layers\(Section[3\.3](https://arxiv.org/html/2608.26149#S3.SS3)\) define fourfunctional layersspecifyingwhatthe system must accomplish in terms of goals and responsibilities\.Modules\(Section[4\.2](https://arxiv.org/html/2608.26149#S4.SS2)\) define fourcomputational modulesspecifyinghoweach layer is implemented through specific algorithms and models\.Steps\(Section[5](https://arxiv.org/html/2608.26149#S5)\) define eightoperational stepsspecifyingwheneach module is executed and how they are orchestrated into an end\-to\-end pipeline\.

Each layer is realized by one primary module, and the eight steps chain these modules together in a complete workflow, adding pre\-processing \(profiling\) and post\-processing \(deployment\) stages\. Figure[1](https://arxiv.org/html/2608.26149#S3.F1)provides an integrated view of these three levels and their correspondences\.

Functional LayersComputational ModulesOperational StepsL1: Semantic AbstractionL2: Relational EngineeringL3: Trans\-Tabular AnalyticsL4: Oper\. OrchestrationM1: Hypergraph Constr\.M2: Temporal EmbeddingsM3: High\-Card\. AttentionM4: Relational DiscoveryS1: Meta\-ProfilingS2: HypergraphS3: PentE Embed\.S4: Contrastive L\.S5: Dyn\. RewiringS6: Causal Infer\.S7: Federated L\.S8: DeploymentFigure 1:Unified view of the framework\.Left:Four functional layers define the system’s goals\.Center:Four computational modules implement these goals\.Right:Eight operational steps orchestrate modules into an end\-to\-end pipeline\. Dashed arrows show layer–module correspondence; solid arrows show which modules are primarily involved in which steps; gray arrows show the sequential execution flow\.
### 3\.3Functional Layers of the Meta\-Model \(MTAM\)

We propose a stratified meta\-model in four functional layers\. Each layer defines a set ofresponsibilitiesthat the system must fulfill; the algorithmic mechanisms fulfilling these responsibilities are described in the corresponding modules \(Section[4\.2](https://arxiv.org/html/2608.26149#S4.SS2)\)\. Critically, these layers are not isolated: they interact through well\-defined data flows \(bottom\-up propagation of learned representations\) and control flows \(top\-down feedback signals that trigger re\-execution of earlier layers when downstream performance degrades or when data drift is detected\)\.

##### Layer 1: Semantic Abstraction\.

This layer transforms raw relational structures into exploitable unified representations\. Its goal is to bridge the gap between the heterogeneous physical schema \(tables, columns, foreign keys\) and a unified semantic space where entities and their attributes can be meaningfully compared and combined\. It encompasses multi\-table knowledge graph construction, relational vector models, and active enriched metadata\. The primary algorithmic realization of this layer is Module 1 \(Hypergraph Construction, Section[4\.2\.1](https://arxiv.org/html/2608.26149#S4.SS2.SSS1)\)\. Layer 1 produces two outputs consumed by downstream layers: the hypergraph structureℋ\\mathcal\{H\}\(consumed by Layers 2 and 3\) and the initial node feature vectorsΦ\\Phi\(consumed by Layer 2 for temporal alignment\)\.

##### Layer 2: Relational Engineering\.

This layer is dedicated to the discovery and optimization of relational structures\. Beyond explicit foreign\-key relationships, it identifies implicit relationships \(e\.g\., columns sharing overlapping value domains\), semantic relationships \(e\.g\., columns referring to the same real\-world concept under different names\), and temporal relationships \(e\.g\., causal ordering of events across tables\)\. It covers automatic relationship discovery, semantic schema alignment, topological graph optimization, and relational metrics computation\. Its primary algorithmic realization is Module 2 \(Multi\-Scale Temporal Embeddings, Section[4\.2\.2](https://arxiv.org/html/2608.26149#S4.SS2.SSS2)\)\. Layer 2 receives the hypergraphℋ\\mathcal\{H\}from Layer 1 and produces enriched temporal embeddings that are passed to Layer 3\. Additionally, Layer 2 sends arefinement signalback to Layer 1 when newly discovered implicit relationships require augmenting the hypergraph with additional hyperedges\.

##### Layer 3: Trans\-Tabular Analytics\.

This layer implements analytical algorithms that exploit the relational structure built by the previous layers to extract cross\-table patterns\. Rather than analyzing each table in isolation, it performs learning and inference over the entire relational graph simultaneously\. It includes relational pattern mining, temporal graph learning, and contextual inference\. Its primary algorithmic realization is Module 3 \(High\-Cardinality Attention, Section[4\.2\.3](https://arxiv.org/html/2608.26149#S4.SS2.SSS3)\)\. Layer 3 receives both the enriched hypergraph from Layer 2 and the PentE embeddings, and produces task\-specific predictions along with attention weights\. These attention weights serve a dual purpose: they feed Layer 4 for monitoring, and they are propagated back to Layer 1 via thedynamic rewiringmechanism \(Step 5\) to adjust hyperedge weights based on their task relevance\.

##### Layer 4: Operational Orchestration\.

This layer manages the lifecycle and governance of the analytical system in production\. It ensures that the models remain calibrated as data evolve, that relational structure changes are detected, and that predictions can be explained\. It covers automated analysis pipelines, relational governance, monitoring, and explainability\. Its primary algorithmic realization is Module 4 \(Differentiable Relational Discovery, Section[4\.2\.4](https://arxiv.org/html/2608.26149#S4.SS2.SSS4)\)\. Layer 4 monitors the outputs of all other layers and issuescontrol signals: when relationship drift is detected, it triggers re\-execution of Layers 1–2; when prediction quality degrades on specific dimensions, it triggers retraining of Layer 3 with adjusted hyperparameters\.

##### Inter\-layer dynamics and architectural properties\.

Figure[2](https://arxiv.org/html/2608.26149#S3.F2)illustrates the complete data flow and control flow between layers\. Three architectural properties ensure the long\-term viability of the system\. First,scalabilityis achieved through the sparse attention mechanism of Layer 3, which bounds the computational cost of cross\-table inference toO​\(n⋅k\)O\(n\\cdot k\)regardless of the total number of entities, and through the differentiable pruning in Layer 1, which keeps the hypergraph size manageable as the database grows\. Second,modularityis ensured by the well\-defined interfaces between layers: each layer communicates through standardized tensor representations \(hypergraph adjacency tensors, embedding matrices, attention weight matrices\), allowing any layer to be replaced or upgraded independently\. Third,extensibilityis supported by the feedback loop architecture: new dimensions of complexity \(e\.g\., spatial data, multimodal inputs\) can be integrated by adding new encoding components to Layer 2 and extending the PentE embedding without modifying the downstream layers\.

Layer 1: Semantic Abstraction\(Module 1\)Layer 2: Relational Engineering\(Module 2\)Layer 3: Trans\-Tabular Analytics\(Module 3\)Layer 4: Operational Orchestration\(Module 4\)Raw schema𝒯\\mathcal\{T\}ℋ,Φ\\mathcal\{H\},\\Phi𝐙PentE,ℋ′\\mathbf\{Z\}\_\{\\text\{PentE\}\},\\mathcal\{H\}^\{\\prime\}y^,𝐀attn\\hat\{y\},\\mathbf\{A\}\_\{\\text\{attn\}\}new hyperedgesrewiring weightsdrift→\\torebuildretrainData flow \(bottom\-up\)Feedback \(refinement\)Control flow \(top\-down\)Figure 2:Functional architecture of the MTAM with inter\-layer interactions\. Blue solid arrows represent bottom\-up data flow \(learned representations propagated from lower to higher layers\)\. Orange dotted arrows represent feedback signals \(Layer 2 augments the hypergraph of Layer 1; Layer 3 rewires hyperedge weights\)\. Red dashed arrows represent control signals from Layer 4, which monitors all layers and triggers re\-execution when drift or degradation is detected\. The standardized interfaces \(ℋ\\mathcal\{H\},𝐙PentE\\mathbf\{Z\}\_\{\\text\{PentE\}\},𝐀attn\\mathbf\{A\}\_\{\\text\{attn\}\}\) between layers ensure modularity and extensibility\.

## 4Relational Hypergraph Transformer Architecture

### 4\.1Motivation and Design Principles

The proposed architecture, calledRelational Hypergraph Transformer\(RHT\), is based on three guiding principles\. The first principle is thegeneralization from graphs to hypergraphs: while standard graph neural networks model binary relationships, relational databases frequently contain n\-ary relationships \(involving more than two entities\), which hypergraphs represent naturally as single hyperedges connecting multiple nodes\. The second principle ismulti\-scale adaptive attention: different dimensions of complexity require specialized attention mechanisms—temporal attention for irregular measurements, relational attention for cross\-table dependencies, and categorical attention for high\-cardinality variables—that must be composed within a unified framework\. The third principle ismodularity and extensibility: each component addresses a specific dimension, enabling independent improvement and flexible composition without requiring architectural redesign\.

### 4\.2Computational Modules

The RHT architecture consists of four computational modules, each implementing a functional layer of the MTAM \(Section[3\.3](https://arxiv.org/html/2608.26149#S3.SS3)\)\. Unlike a simple sequential pipeline, these modules interact through shared data structures and feedback mechanisms, as shown in Figure[3](https://arxiv.org/html/2608.26149#S4.F3)\. Table[2](https://arxiv.org/html/2608.26149#S4.T2)summarizes the input/output specification and functional dependencies of each module\.

Raw tables𝒯\\mathcal\{T\}, metadataM1:HypergraphConstructionM2:TemporalEmbeddingsM3:High\-Card\.AttentionM4:RelationalDiscovery𝒯\\mathcal\{T\}ℋ,Φ\\mathcal\{H\},\\Phiℋ\\mathcal\{H\}ϕtemp\\phi\_\{\\text\{temp\}\}𝐙PentE\\mathbf\{Z\}\_\{\\text\{PentE\}\}y^,𝐀attn\\hat\{y\},\\mathbf\{A\}\_\{\\text\{attn\}\}ℰlatent\\mathcal\{E\}\_\{\\text\{latent\}\}rewiringPredictions, explanations, discovered relationsFigure 3:Inter\-module data flow and feedback in the RHT architecture\. Solid blue arrows indicate primary data flow; dashed orange arrows indicate feedback signals\. Module 4 feeds discovered latent relationships \(ℰlatent\\mathcal\{E\}\_\{\\text\{latent\}\}\) back to Module 1 to augment the hypergraph, and Module 3’s attention weights trigger dynamic rewiring of Module 1’s hyperedge structure\.Table 2:Input/output specification and functional dependencies of each computational module\.ss: sample size per table;dd: embedding dimension;ktk\_\{t\}: number of temporal scales;kk: average hypergraph degree\.

#### 4\.2\.1Module 1: Hypergraph Construction

This module transforms the relational schema into a hypergraph where nodes represent entities \(records\), hyperedges capture n\-ary relationships between tables, and weights reflect the importance and type of each relationship\. Module 1 serves as the entry point of the processing chain: it receives raw tables and metadata from the profiling stage \(Step 1\) and produces the hypergraph structureℋ\\mathcal\{H\}that all downstream modules consume\. Importantly, Module 1 is not invoked only once: it is re\-executed whenever Module 3 triggers dynamic rewiring \(Step 5\) or Module 4 discovers new latent relationships that must be incorporated into the graph\.

Execution scenario \(MIMIC\-IV\)\.Given the 26 MIMIC\-IV tables, Module 1 first detects explicit foreign\-key relationships \(e\.g\.,patients\.subject\_id→\\toadmissions\.subject\_id\), then identifies implicit relationships \(e\.g\., overlapping ICD code sets between thediagnoses\_icdandprocedures\_icdtables\), and constructs hyperedges grouping all entities linked to each hospital admission\. The resulting hypergraph contains approximately5×1055\\times 10^\{5\}nodes and2×1052\\times 10^\{5\}hyperedges, which differentiable pruning reduces to8×1048\\times 10^\{4\}hyperedges while retaining 95% of the relational information \(measured by edge\-cut mutual information\)\.

###### Definition 1\(Relational Hypergraph\)\.

A relational hypergraph is a tupleℋ=\(𝒱,ℰ,𝒲,Φ\)\\mathcal\{H\}=\(\\mathcal\{V\},\\mathcal\{E\},\\mathcal\{W\},\\Phi\)where:

- •𝒱\\mathcal\{V\}is the set of nodes \(entities\)
- •ℰ⊆2𝒱\\mathcal\{E\}\\subseteq 2^\{\\mathcal\{V\}\}is the set of hyperedges
- •𝒲:ℰ→ℝ\+\\mathcal\{W\}:\\mathcal\{E\}\\to\\mathbb\{R\}^\{\+\}assigns weights to hyperedges
- •Φ:𝒱→ℝd\\Phi:\\mathcal\{V\}\\to\\mathbb\{R\}^\{d\}associates feature vectors with nodes

##### Difference between graphs and hypergraphs\.

In a standard graph, each edge connects exactly two nodes\. In a hypergraph, a singlehyperedgecan connect any number of nodes simultaneously\. This is particularly natural for relational databases: a single hospital admission \(one record in the Admissions table\) simultaneously links a patient, a set of diagnoses, a set of prescriptions, and a series of lab measurements\. A hyperedge captures this n\-ary relationship directly, whereas a standard graph would require multiple binary edges and auxiliary nodes, losing the semantics of the group relationship\.

Figure[4](https://arxiv.org/html/2608.26149#S4.F4)illustrates the output of Module 1 on a small medical dataset\.

P1P\_\{1\}P2P\_\{2\}A1A\_\{1\}A2A\_\{2\}D1D\_\{1\}D2D\_\{2\}R​x1Rx\_\{1\}L1L\_\{1\}L2L\_\{2\}Hyperedgee1e\_\{1\}\(Admission 1\)Hyperedgee2e\_\{2\}\(Admission 2\)∙\\bulletPatients \(PP\)∙\\bulletAdmissions \(AA\)∙\\bulletDiagnoses \(DD\)∙\\bulletPrescriptions \(R​xRx\)∙\\bulletLab results \(LL\)Figure 4:Example hypergraph output of Module 1 on a small medical dataset\. Each hyperedge \(dashed or dotted contour\) groups all entities related to a single hospital admission: a patient, an admission record, associated diagnoses, prescriptions, and lab results\. Note that nodeD2D\_\{2\}belongs to both hyperedges, capturing a diagnosis shared across two admissions\.The adaptive construction algorithm is presented below:

Algorithm 1Adaptive Hypergraph Construction0:Set of tables

𝒯=\{T1,…,Tn\}\\mathcal\{T\}=\\\{T\_\{1\},\\ldots,T\_\{n\}\\\}
0:Relational hypergraph

ℋ\\mathcal\{H\}
1:

r​e​l​a​t​i​o​n​s←relations\\leftarrowDetectRelationsMultiLevel

\(𝒯\)\(\\mathcal\{T\}\)
2:

𝒱←\\mathcal\{V\}\\leftarrowExtractEntitiesWithAttributes

\(𝒯\)\(\\mathcal\{T\}\)
3:

ℰ←\\mathcal\{E\}\\leftarrowCreateNaryEdges

\(r​e​l​a​t​i​o​n​s\)\(relations\)
4:

𝒲←\\mathcal\{W\}\\leftarrowComputeRelationSignificance

\(r​e​l​a​t​i​o​n​s\)\(relations\)
5:

ℋc​o​m​p​r​e​s​s​e​d←\\mathcal\{H\}\_\{compressed\}\\leftarrowDifferentiablePruning

\(ℋ\)\(\\mathcal\{H\}\)
6:return

ℋc​o​m​p​r​e​s​s​e​d\\mathcal\{H\}\_\{compressed\}

#### 4\.2\.2Module 2: Multi\-Scale Temporal Embeddings

To handle repeated measurements at variable frequencies \(dimension 5\), this module uses an adaptive temporal embedding function inspired by Time2Vec\[[10](https://arxiv.org/html/2608.26149#bib.bib10)\]\. Module 2 receives the hypergraphℋ\\mathcal\{H\}and node featuresΦ\\Phifrom Module 1, and produces the PentE embeddings𝐙\\mathbf\{Z\}that integrate all five dimensions into a unified latent space\. Its output is consumed both by Module 3 \(for cross\-table attention\) and by Module 4 \(for temporal drift detection\)\.

Execution scenario \(MIMIC\-IV\)\.Consider a patient with laboratory measurements recorded at irregular intervals \(e\.g\., blood glucose at hours 0, 2, 7, 24, 48 post\-admission\)\. Module 2 encodes each timestamp using multi\-resolution sinusoidal functions, with learned frequenciesωi\\omega\_\{i\}that automatically adapt to capture both short\-term variations \(hourly fluctuations\) and long\-term trends \(daily cycles\)\. The temporal embeddings are then concatenated with semantic, relational, categorical, and volumetric components to form the PentE vector for each measurement event\.

###### Definition 2\(Multi\-Resolution Temporal Embedding\)\.

For a timestamptt, the temporal embedding is defined by:

ϕtemp​\(t\)=\[ω0​t,sin⁡\(ω1​t\),sin⁡\(ω2​t\),…,sin⁡\(ωk​t\)\]\\phi\_\{\\text\{temp\}\}\(t\)=\\left\[\\omega\_\{0\}t,\\sin\(\\omega\_\{1\}t\),\\sin\(\\omega\_\{2\}t\),\\ldots,\\sin\(\\omega\_\{k\}t\)\\right\]\(1\)where the frequencies\{ωi\}i=1k\\\{\\omega\_\{i\}\\\}\_\{i=1\}^\{k\}are learned to capture different temporal scales\.

Cross\-table alignment of irregular time series is achieved via a temporal attention layer\.

#### 4\.2\.3Module 3: High\-Cardinality Attention

To efficiently process high cardinality categorical variables \(dimension 3\), this module implements asparse attentionmechanism guided by relational structure\. Module 3 receives the hypergraphℋ\\mathcal\{H\}from Module 1 and the PentE embeddings𝐙\\mathbf\{Z\}from Module 2, and produces task\-specific predictionsy^\\hat\{y\}along with attention weight matrices𝐀attn\\mathbf\{A\}\_\{\\text\{attn\}\}\. The attention weights are the key interface to other modules: they are consumed by Module 4 for monitoring and explanation, and they are fed back to Module 1 via the dynamic rewiring mechanism to adjust hyperedge importance based on task relevance\.

The key design goal is to reduce the computational complexity of standard attention fromO​\(n2\)O\(n^\{2\}\)toO​\(n⋅k\)O\(n\\cdot k\), wherekkis the average degree of the relational graph,without sacrificing recall on rare categories or rare relationships\. This is achieved by restricting attention computation to pairs of entities that are relationally connected in the hypergraph, thereby filtering out the vast majority of irrelevant pairs while preserving all existing relational signals—including those involving rare categories\.

Execution scenario \(MIMIC\-IV\)\.For ICD\-10 code prediction, Module 3 computes attention between a patient’s admission embedding and all diagnosis codes reachable through the hypergraph \(typicallyk=15k=15–5050codes per admission, versusn\>10,000n\>10\{,\}000total codes\)\. The memory bank stores prototypes for rare codes \(frequency<0\.1%<0\.1\\%, approximately 7,000 of the 10,000\+ ICD codes\), enabling few\-shot recognition\. The attention weights reveal which inter\-table relationships \(e\.g\., specific lab result patterns linked to specific diagnoses\) are most informative for each prediction\.

###### Definition 3\(Sparse Relational Attention\)\.

Sparse relational attention is defined by:

Attention​\(Q,K,V\)=softmax​\(Q​KTdk\+Mmask\)​V\\text\{Attention\}\(Q,K,V\)=\\text\{softmax\}\\left\(\\frac\{QK^\{T\}\}\{\\sqrt\{d\_\{k\}\}\}\+M\_\{\\text\{mask\}\}\\right\)V\(2\)whereMmask∈\{0,−∞\}n×nM\_\{\\text\{mask\}\}\\in\\\{0,\-\\infty\\\}^\{n\\times n\}is a learned mask matrix that enforces sparsity according to relational structure:

Mmask​\[i,j\]=\{0if​i​and​j​are relationally connected−∞otherwiseM\_\{\\text\{mask\}\}\[i,j\]=\\begin\{cases\}0&\\text\{if \}i\\text\{ and \}j\\text\{ are relationally connected\}\\\\ \-\\infty&\\text\{otherwise\}\\end\{cases\}\(3\)

Amemory bankfor few\-shot learning stores prototypes of rare categories, enabling their efficient recognition even when training examples are scarce\.

#### 4\.2\.4Module 4: Differentiable Relational Discovery

This module learns end\-to\-end to discover latent relationships between tables that are not explicitly defined in the schema\. Module 4 is the only module that depends on all three preceding modules: it receives the current hypergraphℋ\\mathcal\{H\}from Module 1, the temporal embeddings from Module 2, and the predictions and attention weights from Module 3\. Its outputs are twofold: newly discovered latent relationshipsℰlatent\\mathcal\{E\}\_\{\\text\{latent\}\}, which are fed back to Module 1 to augment the hypergraph, and monitoring signals \(drift alerts, explanation dashboards\) consumed by the end user and the operational pipeline\.

Execution scenario \(MIMIC\-IV\)\.Module 4 discovers that certain pairs of tables not linked by explicit foreign keys nonetheless exhibit strong statistical dependencies—for example, that specific patterns in thelabeventstable \(rising creatinine levels\) are predictive of entries in theprescriptionstable \(initiation of renal\-protective medications\), mediated by latent clinical reasoning not captured in the schema\. These discovered relationships are represented as soft hyperedges with learned weights and are injected back into the hypergraph for the next training iteration, progressively enriching the relational structure\.

###### Definition 4\(Relational Discovery Loss\)\.

The loss function for relational discovery combines three terms:

ℒrel=α​ℒtask\+β​ℒsparse\+γ​ℒsemantic\\mathcal\{L\}\_\{\\text\{rel\}\}=\\alpha\\mathcal\{L\}\_\{\\text\{task\}\}\+\\beta\\mathcal\{L\}\_\{\\text\{sparse\}\}\+\\gamma\\mathcal\{L\}\_\{\\text\{semantic\}\}\(4\)where:

- •ℒtask\\mathcal\{L\}\_\{\\text\{task\}\}is the main supervised task loss
- •ℒsparse\\mathcal\{L\}\_\{\\text\{sparse\}\}penalizes excessive graph density
- •ℒsemantic\\mathcal\{L\}\_\{\\text\{semantic\}\}encourages semantic coherence of discovered relationships

## 5Eight\-Step Methodology

Our multi\-table analysis methodology is articulated in eight sequential steps, with feedback loops enabling iterative adjustment\. These steps orchestrate the four computational modules \(Section[4\.2](https://arxiv.org/html/2608.26149#S4.SS2)\) into a complete end\-to\-end pipeline: Step 1 is a preliminary profiling stage; Steps 2–3 invoke Module 1; Steps 4–5 invoke Modules 2 and 3; Steps 6–7 extend the framework with causal inference and federated learning; Step 8 handles deployment and monitoring via Module 4\. See Figure[1](https://arxiv.org/html/2608.26149#S3.F1)for the full mapping\.

### 5\.1Step 1: Predictive Meta\-Profiling

Before any analysis, this preliminary step automatically characterizes data according to the five dimensions\. It serves as the entry point of the pipeline and determines the optimal configuration for all subsequent steps\. The profiling process pursues four objectives: inferring the complete relational schema through few\-shot learning on small data samples, predicting probable relationship types between tables, estimating cardinality and distribution of categorical variables, and recommending the optimal analytical pipeline through a relational AutoML strategy\.

Method:Schema inference uses lightweight sampling\-based profiling: a small sample of rows \(typically 1% or 10,000 rows, whichever is smaller\) is drawn from each table\. Column\-type detection, cardinality estimation \(via HyperLogLog\), and foreign\-key candidate identification \(via value\-overlap statistics\) are performed on these samples\. The resulting dimensional profile is fed into a rule\-based recommender that selects appropriate hyperparameters for the downstream modules\.

The step produces three outputs: a quantitative dimensional profile characterizing the data along the five dimensions, a table meta\-graph capturing the inferred relational structure, and a recommended pipeline configuration that parameterizes all subsequent steps\.

### 5\.2Step 2: Automated Hypergraph Construction

Transformation of relational schema into hypergraph according to Algorithm[1](https://arxiv.org/html/2608.26149#alg1)\. This step is the primary invocation ofModule 1\(Section[4\.2\.1](https://arxiv.org/html/2608.26149#S4.SS2.SSS1)\)\.

Process:The transformation proceeds in three stages: multi\-level relationship detection \(covering explicit foreign keys, implicit value overlaps, semantic correspondences, and temporal dependencies\), followed by weighted hyperedge construction that assigns importance scores to each detected relationship, and finally differential compression for scalability that prunes low\-significance hyperedges while preserving gradient flow for end\-to\-end training\.

### 5\.3Step 3: Unified 5D Embedding \(PentE\)

Creation of a unified latent space integrating the five dimensions\. This step combines the output ofModule 1\(hypergraph structure\) with the encoders fromModule 2\(temporal embeddings\) andModule 3\(categorical embeddings\)\.

###### Definition 5\(PentE \- Pentadimensional Embedding\)\.

For an entityee, its PentE embedding is defined by:

𝐳e=ϕsem​\(e\)⊕ϕrel​\(e\)⊕ϕtemp​\(e\)⊕ϕcat​\(e\)⊕ϕvol​\(e\)\\mathbf\{z\}\_\{e\}=\\phi\_\{\\text\{sem\}\}\(e\)\\oplus\\phi\_\{\\text\{rel\}\}\(e\)\\oplus\\phi\_\{\\text\{temp\}\}\(e\)\\oplus\\phi\_\{\\text\{cat\}\}\(e\)\\oplus\\phi\_\{\\text\{vol\}\}\(e\)\(5\)where⊕\\oplusrepresents concatenation and eachϕ∗\\phi\_\{\*\}encodes a specific dimension:

- •ϕsem\\phi\_\{\\text\{sem\}\}: Semantic embedding of attributes
- •ϕrel\\phi\_\{\\text\{rel\}\}: Embedding of position in relational graph \(via Module 1\)
- •ϕtemp\\phi\_\{\\text\{temp\}\}: Multi\-resolution temporal embedding \(via Module 2, Section[4\.2\.2](https://arxiv.org/html/2608.26149#S4.SS2.SSS2)\)
- •ϕcat\\phi\_\{\\text\{cat\}\}: Hierarchical embedding of categories \(via Module 3, Section[4\.2\.3](https://arxiv.org/html/2608.26149#S4.SS2.SSS3)\)
- •ϕvol\\phi\_\{\\text\{vol\}\}: Volumetric normalization embedding

The PentE space is constrained by three regularizations that ensure geometric coherence of the embedding space:relational preservation, which enforces that linked entities must be close in the embedding space;temporal continuity, which promotes smooth evolution of temporal embeddings to reflect the continuous nature of physical processes; andcategorical similarity, which ensures that semantically close categories \(e\.g\., related ICD\-10 codes within the same chapter\) have similar embeddings\.

### 5\.4Step 4: Relational Contrastive Learning

Model training via a specialized contrastive loss function\. This step trains the encoders ofModules 2 and 3using the relational structure from Module 1\.

###### Definition 6\(Relational\-Temporal Contrastive Loss\)\.

ℒCRT=α​ℒrel\+β​ℒtemp\+γ​ℒsem\\mathcal\{L\}\_\{\\text\{CRT\}\}=\\alpha\\mathcal\{L\}\_\{\\text\{rel\}\}\+\\beta\\mathcal\{L\}\_\{\\text\{temp\}\}\+\\gamma\\mathcal\{L\}\_\{\\text\{sem\}\}\(6\)where:

- •ℒrel\\mathcal\{L\}\_\{\\text\{rel\}\}brings together linked entities and separates unlinked ones
- •ℒtemp\\mathcal\{L\}\_\{\\text\{temp\}\}aligns temporally close entities
- •ℒsem\\mathcal\{L\}\_\{\\text\{sem\}\}encourages semantic coherence

### 5\.5Step 5: Dynamic Graph Rewiring

Attention mechanism allowing the model to dynamically reweight the relational graph based on the task\. This step refines the hypergraph structure produced by Module 1 using the trained attention weights fromModule 3\(Section[4\.2\.3](https://arxiv.org/html/2608.26149#S4.SS2.SSS3)\)\.

Principle:The model learns to identify important relationships for a specific task and adjusts hyperedge weights accordingly\.

Implementation:Attention layer on hyperedges with gating mechanism\.

### 5\.6Step 6: Relational Causal Inference

Framework extension to discover causal relationships across tables\. This step builds upon the refined relational structure from Steps 4–5 and adds a causal analysis layer\. The methodology combines three complementary approaches: Double Machine Learning for robust causal effect estimation in the presence of high\-dimensional confounders, relational counterfactual reasoning that addresses questions of the form“What would happen if this relationship didn’t exist?”, and Granger causality tests adapted to the graph structure to identify temporal causal dependencies across tables\.

### 5\.7Step 7: Federated Multi\-Table Learning

For distributed or sensitive data scenarios \(e\.g\., healthcare\), privacy\-preserving federated learning\. This step enables training across institutional boundaries without sharing raw data\. The strategy proceeds in three stages: local learning on schema fragments at each participating site, secure aggregation of relational embeddings using cryptographic protocols that prevent the reconstruction of individual data points, and the application of differential privacy at the relationship level to provide formal privacy guarantees while preserving the relational structure needed for effective cross\-table learning\.

### 5\.8Step 8: Operationalization and Monitoring

Production deployment with continuous monitoring of relational quality\. This step is the primary invocation ofModule 4\(Section[4\.2\.4](https://arxiv.org/html/2608.26149#S4.SS2.SSS4)\), which detects relationship drift and triggers incremental retraining when the relational structure evolves\. The deployment infrastructure comprises four integrated components: a relational MLOps pipeline that automates the transition from trained model to production service, a relationship drift detection system that monitors changes in the statistical properties of inter\-table relationships over time, an incremental retraining mechanism that updates the model in response to detected drift without requiring full retraining from scratch, and an explainability dashboard that provides domain experts with interpretable visualizations of the hypergraph structure and attention\-based feature attributions\.

## 6Key Technical Innovations

The previous sections described the architecture \(modules\) and the methodology \(steps\)\. This section zooms in on specificalgorithmic innovationsthat cut across multiple modules and constitute the main technical contributions of this work\. These innovations are presented separately because each one involves design choices and implementation details that affect several modules simultaneously\.

### 6\.1Sparse Relational Attention

As defined in Module 3 \(Section[4\.2\.3](https://arxiv.org/html/2608.26149#S4.SS2.SSS3)\), this mechanism reduces attention computational complexity fromO​\(n2\)O\(n^\{2\}\)toO​\(n⋅k\)O\(n\\cdot k\)wherekkis the average degree in the relational graph\. The formal procedure is presented in Algorithm[2](https://arxiv.org/html/2608.26149#alg2)\.

Algorithm 2Sparse Relational Attention0:Query matrix

Q∈ℝn×dkQ\\in\\mathbb\{R\}^\{n\\times d\_\{k\}\}, Key matrix

K∈ℝn×dkK\\in\\mathbb\{R\}^\{n\\times d\_\{k\}\}, Value matrix

V∈ℝn×dvV\\in\\mathbb\{R\}^\{n\\times d\_\{v\}\}, Hypergraph adjacency

𝒜∈\{0,1\}n×n\\mathcal\{A\}\\in\\\{0,1\\\}^\{n\\times n\}
0:Contextualized representations

O∈ℝn×dvO\\in\\mathbb\{R\}^\{n\\times d\_\{v\}\}, Attention weights

W∈ℝn×nW\\in\\mathbb\{R\}^\{n\\times n\}
1:// Step 1: Compute raw attention scores \(restricted to relational neighbors\)

2:foreach entity

i∈\{1,…,n\}i\\in\\\{1,\\ldots,n\\\}do

3:

𝒩​\(i\)←\{j:𝒜​\[i,j\]=1\}\\mathcal\{N\}\(i\)\\leftarrow\\\{j:\\mathcal\{A\}\[i,j\]=1\\\}⊳\\trianglerightRelational neighborhood ofii

4:foreach

j∈𝒩​\(i\)j\\in\\mathcal\{N\}\(i\)do

5:

S​\[i,j\]←Qi⋅Kj⊤dkS\[i,j\]\\leftarrow\\frac\{Q\_\{i\}\\cdot K\_\{j\}^\{\\top\}\}\{\\sqrt\{d\_\{k\}\}\}⊳\\trianglerightScaled dot\-product score

6:endfor

7:endfor

8:// Step 2: Construct sparse mask from relational structure

9:foreach pair

\(i,j\)∈\{1,…,n\}2\(i,j\)\\in\\\{1,\\ldots,n\\\}^\{2\}do

10:

M​\[i,j\]←\{0if​𝒜​\[i,j\]=1−∞otherwiseM\[i,j\]\\leftarrow\\begin\{cases\}0&\\text\{if \}\\mathcal\{A\}\[i,j\]=1\\\\ \-\\infty&\\text\{otherwise\}\\end\{cases\}
11:endfor

12:// Step 3: Masked softmax normalization

13:foreach entity

i∈\{1,…,n\}i\\in\\\{1,\\ldots,n\\\}do

14:

W​\[i,:\]←softmax​\(S​\[i,:\]\+M​\[i,:\]\)W\[i,:\]\\leftarrow\\text\{softmax\}\(S\[i,:\]\+M\[i,:\]\)⊳\\trianglerightOnly neighbors receive non\-zero weight

15:endfor

16:// Step 4: Weighted aggregation

17:

O←W⋅VO\\leftarrow W\\cdot V⊳\\trianglerightOutput:Oi=∑j∈𝒩​\(i\)W​\[i,j\]⋅VjO\_\{i\}=\\sum\_\{j\\in\\mathcal\{N\}\(i\)\}W\[i,j\]\\cdot V\_\{j\}

18:return

O,WO,W

###### Proposition 1\(Complexity of sparse relational attention\)\.

Letℋ\\mathcal\{H\}be a relational graph withnnnodes whose neighbourhoods are represented as an edge list of total sizeE=∑i\|𝒩​\(i\)\|E=\\sum\_\{i\}\|\\mathcal\{N\}\(i\)\|, with average degreek=E/nk=E/n\. Computing relational attention over this edge list \(Algorithm[2](https://arxiv.org/html/2608.26149#alg2), edge\-list form\) requiresO​\(E⋅dk\)=O​\(n⋅k⋅dk\)O\(E\\cdot d\_\{k\}\)=O\(n\\cdot k\\cdot d\_\{k\}\)time andO​\(E\)=O​\(n⋅k\)O\(E\)=O\(n\\cdot k\)space, versusO​\(n2⋅dk\)O\(n^\{2\}\\cdot d\_\{k\}\)time andO​\(n2\)O\(n^\{2\}\)space for dense attention\. Whenk≪nk\\ll n, the asymptotic reduction factor isn/kn/k\.

###### Proof sketch\.

Each edge\(i,j\)\(i,j\)contributes one scaled dot product \(O​\(dk\)O\(d\_\{k\}\)\), one segment\-softmax term, and one weighted value aggregation; there areEEedges and no operation ranges over non\-adjacent pairs\. Non×nn\\times ntensor is materialised\. The dense baseline forms alln2n^\{2\}scores explicitly\.□\\square∎

Implementation note\.The edge\-list \(COO\) implementation in our released code realises this bound and is verified to produce attention outputs numerically identical \(max\. absolute difference<10−7<10^\{\-7\}\) to the dense masked formulation on random graphs, confirming that the complexity reduction incurs no change in the computed function\. The reference implementation in the repository additionally provides the dense masked version of Definition[6\.1](https://arxiv.org/html/2608.26149#S6.SS1)for pedagogical clarity\.

### 6\.2Hierarchical High\-Cardinality Encoding

For very high cardinality categorical variables \(e\.g\., ICD\-10 codes\), we propose a three\-level hierarchical encoding\. This innovation is used by Module 3 and contributes to the categorical component \(ϕcat\\phi\_\{\\text\{cat\}\}\) of the PentE embedding \(Step 3\)\.

Process:The encoding proceeds in three stages\. Theinitial embeddingstage generates base representations using either pre\-trained embeddings \(Word2Vec on textual descriptions of categories\) or embeddings learned from co\-occurrence patterns in the data\. Next,hierarchical clusteringconstructs a decision tree in embedding space that groups semantically similar categories into a hierarchy\. Finally, thecomposite embeddingfor a categoryccis formed by concatenating three components:

𝐞c=𝐞code⊕𝐞cluster⊕𝐞parent\\mathbf\{e\}\_\{c\}=\\mathbf\{e\}\_\{\\text\{code\}\}\\oplus\\mathbf\{e\}\_\{\\text\{cluster\}\}\\oplus\\mathbf\{e\}\_\{\\text\{parent\}\}\(7\)
Advantage:Rare categories inherit information from their cluster and parent in the hierarchy, enabling efficient generalization\.

### 6\.3Temporal\-Relational Message Passing

Extension of GNN message passing for temporal graphs with irregular measurements\. This innovation is the core mechanism of Module 2 and is invoked during Steps 3 and 4\.

###### Definition 7\(Temporal\-Relational Message Passing\)\.

The update of a nodevv’s representation at timettis defined by:

𝐡v\(t\+1\)=ϕ​\(𝐡vt,⨁\(u,τ\)∈𝒩​\(v\)ψ​\(𝐡uτ,𝐞u​v,g​\(t−τ\)\)\)\\mathbf\{h\}\_\{v\}^\{\(t\+1\)\}=\\phi\\left\(\\mathbf\{h\}\_\{v\}^\{t\},\\bigoplus\_\{\(u,\\tau\)\\in\\mathcal\{N\}\(v\)\}\\psi\\left\(\\mathbf\{h\}\_\{u\}^\{\\tau\},\\mathbf\{e\}\_\{uv\},g\(t\-\\tau\)\\right\)\\right\)\(8\)where:

- •𝒩​\(v\)\\mathcal\{N\}\(v\)is the temporal neighborhood ofvv\(nodes connected at different times\)
- •𝐞u​v\\mathbf\{e\}\_\{uv\}is the embedding of the relationship betweenuuandvv
- •g​\(t−τ\)g\(t\-\\tau\)is a learned temporal decay function
- •ϕ\\phiandψ\\psiare neural networks
- •⨁\\bigoplusis temporal aggregation \(e\.g\., weighted attention\)

### 6\.4Causal Relational Discovery

Method for identifying latent causal relationships in multi\-table data\. This innovation is invoked during Step 6\.

Approach:The method proceeds in three stages: Granger causality tests adapted to graph\-structured data to identify directed temporal dependencies, structural equation modeling \(SEM\) on the hypergraph to formalize the causal mechanisms underlying observed correlations, and validation through do\-calculus and counterfactual analysis to distinguish genuine causal relationships from spurious associations induced by confounding variables\.

Medical example:Discovery that medication administration \(Prescriptions table\) causes specific variation in biological measurements \(Laboratory table\), mediated by certain diagnoses \(Diagnostics table\)\.

## 7Comparative Benchmark and Evaluation

### 7\.1Reference Datasets

For rigorous evaluation, we propose using the following datasets:

#### 7\.1\.1MIMIC\-IV \(Medical Information Mart for Intensive Care\)

MIMIC\-IV constitutes the primary evaluation dataset due to its comprehensive coverage of all five dimensions: it comprises 26 interconnected tables with complex relationships \(patients→\\toadmissions→\\todiagnoses→\\toprescriptions→\\tomeasurements\), over 15 million repeated measurements at irregular intervals, and over 10,000 unique ICD codes representing high cardinality with heavily imbalanced distributions\. The evaluation tasks on MIMIC\-IV include hospital mortality prediction, length of stay prediction, multi\-label diagnosis classification, and adverse event detection, each exercising different combinations of the five dimensions\.

#### 7\.1\.2Amazon Multi\-Table Dataset

The Amazon dataset provides a complementary evaluation context with emphasis on many\-to\-many relationships and hierarchical categorization\. It comprises five interconnected tables \(Products, Reviews, Users, Metadata, Categories\) linked by complex many\-to\-many relationships, with hierarchical product categories exhibiting high cardinality and interaction time series capturing user behavior over time\.

#### 7\.1\.3Financial Transactions Dataset

The financial transactions dataset tests the framework under extreme temporal resolution and dynamic relational structure\. It consists of multi\-institutional banking transactions with fine temporal relationships at millisecond granularity, high cardinality of transaction codes, and a dynamic relational graph whose structure evolves as new accounts and transaction patterns emerge over time\.

### 7\.2Multidimensional Evaluation Metrics

We propose a comprehensive set of metrics structured by dimension\. For each metric, we specify the evaluation protocol \(what is measured, against what ground truth, and under what conditions\)\.

#### 7\.2\.1Dimension 1: Volume \(Scalability\)

Scalability is assessed through three complementary measures\.Throughputquantifies the number of relationships processed per second on a fixed hardware configuration \(to be specified in the experimental setup\)\. Thescalability curvetraces wall\-clock training time as a function of the number of tables, measured by progressively adding tables fromn=1n=1ton=100n=100on subsets of the MIMIC\-IV dataset, testing how computation time scales with relational complexity\. Thecompression ratio, defined as the original data size in bytes divided by the hypergraph representation size, measures the efficiency of the structural encoding\.

#### 7\.2\.2Dimension 2: Many Variables

The handling of high\-dimensional feature spaces is evaluated through two metrics\.Feature importance consistencymeasures the stability of feature selection across 10 bootstrap samples using the Jaccard similarity of the top\-kkselected features, ensuring that the model’s reliance on specific features is robust rather than an artifact of sampling variability\. Thedimensional preservation scorecomputes the mutual informationI​\(X;Z\)I\(X;Z\)between the original feature matrixXXand the reduced PentE representationZZ, normalized byH​\(X\)H\(X\), quantifying how much information from the original feature space is retained in the compressed representation\.

#### 7\.2\.3Dimension 3: High Cardinality

High cardinality handling is evaluated through two metrics that capture both predictive and representational quality\. TheRare Category Recall@kk\(RCR@kk\) measures, for a downstream classification task \(e\.g\., ICD code prediction on MIMIC\-IV\), the recall computed on the subset of categories with training\-set frequency below 0\.1%\. Specifically, among test samples belonging to rare categories, it computes the fraction for which the correct category appears in the model’s top\-kkpredictions \(formal definition in Equation[9](https://arxiv.org/html/2608.26149#S7.E9)\)\. Thesemantic coherencemetric evaluates the quality of the learned categorical embedding space by computing the ratio of average intra\-cluster cosine similarity to average inter\-cluster cosine similarity, where clusters are defined by the ICD\-10 chapter hierarchy\. A high ratio indicates that the embedding space respects the known semantic structure of the category system\.

RCR​@​k=1\|𝒞rare\|​∑c∈𝒞rare𝟙​\[rank​\(c\)≤k\]\\text\{RCR\}@k=\\frac\{1\}\{\|\\mathcal\{C\}\_\{\\text\{rare\}\}\|\}\\sum\_\{c\\in\\mathcal\{C\}\_\{\\text\{rare\}\}\}\\mathbb\{1\}\[\\text\{rank\}\(c\)\\leq k\]\(9\)where𝒞rare=\{c:freq​\(c\)<0\.001\}\\mathcal\{C\}\_\{\\text\{rare\}\}=\\\{c:\\text\{freq\}\(c\)<0\.001\\\}is the set of categories with training\-set frequency below 0\.1%\.

#### 7\.2\.4Dimension 4: Multiple Tables

Relational discovery quality is assessed through two complementary measures\. Therelational discovery F1\-scoreevaluates the precision and recall of latent relationships inferred by Module 4 against a held\-out ground truth constructed as follows: from the full database schema, a random subset of known foreign\-key relationships is hidden from the model during training, and the model’s ability to rediscover them is measured\. Additionally, domain experts annotate a set of known semantic relationships \(e\.g\., shared clinical concepts across tables\) that serve as complementary ground truth\. Theschema completion accuracyreports the fraction of hidden relationships correctly recovered through this protocol\.

#### 7\.2\.5Dimension 5: Repeated Measurements

Temporal modeling quality is evaluated through two metrics\. Theirregular time series imputation erroris assessed via a masking protocol: for each time series, 20% of observed values are randomly masked, and the model predicts them\. For quantitative features, Mean Absolute Error \(MAE\) is reported, preferred over MSE for its interpretability in clinical units and robustness to outliers in medical data\. For categorical features measured over time \(e\.g\., repeated diagnostic codes\), accuracy and macro\-F1 score on the masked values are reported\. This dual protocol ensures coverage of both quantitative and qualitative temporal variables\. Thetemporal pattern discovery ratemeasures the ratio of statistically significant temporal patterns identified \(tested via permutation tests atp<0\.01p<0\.01\) to the number of expert\-validated patterns in the dataset\.

#### 7\.2\.6Anomaly and Extreme Value Detection

In addition to the dimension\-specific metrics above, we evaluate the model’s ability to detect extreme or anomalous values, which is critical in healthcare \(e\.g\., detecting abnormal lab results\) and finance \(e\.g\., detecting fraudulent transactions\)\. Theextreme value detection precision and recallare computed using domain\-specific thresholds \(e\.g\., clinically defined critical lab value ranges from MIMIC\-IV documentation\) to measure the model’s ability to flag extreme values\. Therelational anomaly detection AUROCquantifies the area under the ROC curve for detecting structurally anomalous relationships \(e\.g\., a prescription inconsistent with a patient’s diagnosis history\), evaluated against expert annotations\.

#### 7\.2\.7Holistic Metrics

###### Definition 8\(5D Integration Score\)\.

S5​D=5∑i=151siS\_\{5D\}=\\frac\{5\}\{\\sum\_\{i=1\}^\{5\}\\frac\{1\}\{s\_\{i\}\}\}\(10\)where eachsi∈\[0,1\]s\_\{i\}\\in\[0,1\]is the normalized score for dimensionii:

si=scorei−minimaxi−minis\_\{i\}=\\frac\{\\text\{score\}\_\{i\}\-\\text\{min\}\_\{i\}\}\{\\text\{max\}\_\{i\}\-\\text\{min\}\_\{i\}\}\(11\)The harmonic mean penalizes low performance on any single dimension, reflecting the principle that a holistic solution must perform well across all five dimensions simultaneously\.

###### Definition 9\(Multi\-Table Information Gain\)\.

GM​T=I​\(post\-fusion\)I​\(pre\-fusion\)G\_\{MT\}=\\frac\{I\(\\text\{post\-fusion\}\)\}\{I\(\\text\{pre\-fusion\}\)\}\(12\)whereI​\(⋅\)I\(\\cdot\)measures mutual information between features and target\.

### 7\.3Experimental Setup

##### Dataset\.

We evaluate onSynthea\[[5](https://arxiv.org/html/2608.26149#bib.bib5)\], an open generator of synthetic but statistically realistic electronic health records, requiring no data\-use agreement\. We export thepatients,encounters, andconditionstables using Synthea’s default SNOMED CT coding \(system URI:[http://snomed\.info/sct](http://snomed.info/sct)\)\. We generate 22,913 synthetic patients, yielding 1,320,007 encounters and 821,371 condition records spanning 313 distinct SNOMED condition codes\. The code\-frequency distribution is long\-tailed: 77\.32% of codes occur in fewer than0\.1%0\.1\\%of encounters and constitute the rare\-category set𝒞rare\\mathcal\{C\}\_\{\\text\{rare\}\}\.

##### Task\.

Given a patient’s demographic features and an encounter’s attributes and relational context, predict the multi\-label set of SNOMED condition codes assigned at that encounter\. This task isolates Dimension 3 \(high cardinality\) and Dimension 4 \(relational context\), which is where our architecture is designed to help\.

##### Splits\.

All splits are performed*at the patient level*so that no patient’s encounters appear in more than one split \(preventing leakage\), with a70/15/1570/15/15train/validation/test partition\. We report mean±\\pmstandard deviation over 5 random seeds\.

##### Baselines\.

We compare against: \(i\)SQL Wide Table \+ XGBoost, a tabular baseline on flattened patient\+\+encounter features \(no relational structure\); \(ii\)GraphSAGE, a relational mean\-aggregation encoder that models inter\-table structure but treats all categories uniformly \(no hierarchy\); and \(iii\)TGN\[[7](https://arxiv.org/html/2608.26149#bib.bib7)\], a temporal graph network\. For TGN we report published relational results and keep our harness integration as a clearly labelled stub \(to avoid claiming unexecuted numbers\)\.

##### Metrics\.

We reportRCR@kk\(rare\-category recall@kk, Eq\.[9](https://arxiv.org/html/2608.26149#S7.E9)\) fork∈\{10,50\}k\\in\\\{10,50\\\}, restricted to𝒞rare\\mathcal\{C\}\_\{\\text\{rare\}\};macro\-F1over all codes; andsemantic coherence, the ratio of intra\- to inter\-cluster cosine similarity of the learned code embeddings, with clusters defined by first\-character prefix groups of the SNOMED codes \(a structural grouping analogous to ICD\-10 chapter\-level hierarchy\)\. Coherence is computed with a numerically stabilised estimator to avoid the small\-sample blow\-up of the naive ratio\.

##### Implementation\.

All models use embedding dimension 64, 4 attention heads, and are trained for 50 epochs with Adam \(lr 1e\-3\) on an NVIDIA GB10 GPU \(CUDA\)\. Full configurations are released with the code\.

### 7\.4Results

Table[3](https://arxiv.org/html/2608.26149#S7.T3)reports the main comparison using the values emitted byexperiments/run\.py\(results\.tex\), aggregated as mean±\\pmstd over 5 seeds\.

Table 3:Main results on Synthea \(multi\-label condition\-code prediction, SNOMED CT, 5 seeds×\\times50 epochs\)\. Mean±\\pmstd over 5 seeds\. Higher is better for all metrics\.##### Discussion\.

Table[3](https://arxiv.org/html/2608.26149#S7.T3)reports measured results on 22,913 Synthea patients\. XGBoost achieves the highest raw RCR@10/50 and macro\-F1, which is expected given its strong inductive bias for tabular data and the moderate dataset size\. GraphSAGE demonstrates non\-trivial relational recall \(RCR@10 = 0\.082\) but lacks semantic coherence\. RHT \(full\) is the only model producing meaningful semantic coherence \(1\.52±\\pm0\.03\), confirming that hierarchical categorical encoding \(M3\) and the relational graph structure jointly organise predictions into coherent SNOMED code groups\. The ablation in Table[4](https://arxiv.org/html/2608.26149#S7.T4)shows that disabling M3 collapses coherence to 1\.00±\\pm0\.00 \(random baseline\), confirming M3 as the decisive driver of semantic structure\. Macro\-F1 remains low across all models, consistent with the well\-known class\-imbalance challenge of multi\-label condition coding on synthetic data with 313 distinct codes\.

### 7\.5Ablation

To isolate each module’s contribution, we retrain RHT with one module disabled at a time \(M1: relational attention; M2: temporal embedding; M3: hierarchical categorical encoding; M4: relational\-discovery auxiliary loss\)\. The released harness implements these as configuration switches\. Table[4](https://arxiv.org/html/2608.26149#S7.T4)reports the resulting metrics from the same 5\-seed run\.

Table 4:Module ablation on Synthea \(5 seeds×\\times50 epochs\)\. Each row disables one module\. RHT \(full\) is the complete model; w/o Mx disables module Mx\.
### 7\.6Phase 1 Pilot on Existing Benchmark Data \(MIMIC\-IV Demo\)

To concretely instantiate the Phase 1 roadmap on an*existing benchmark*, we executed the local MIMIC\-IV demo benchmark harness \(benchmarks/local\_csv\_benchmark\.py\) on all subset scales \(QUARTER/HALF/FULL\), for both RHT configurations \(TinyEmbed and Full\), and over 5 random seeds\. Table[5](https://arxiv.org/html/2608.26149#S7.T5)reports measured throughput/compression/runtime/memory statistics as mean±\\pmstd\.

Table 5:Phase 1 pilot results on local MIMIC\-IV demo CSV benchmark \(5 seeds, 5 epochs\)\. This table reports measured scalability/resource metrics used to replace purely anticipated claims in the validation plan\.These pilot measurements confirm the expected scaling trend of sparse relational processing on progressively larger subsets, with throughput reaching≈1\.01×105\\approx 1\.01\\times 10^\{5\}rows/s and compression ratio reaching≈4\.09\\approx 4\.09on the FULL subset\. In contrast, D3/D4/D5 quality metrics remain weak or unstable in this demo setting \(rare\-category recall and relation\-discovery near zero; temporal imputation requiring stronger calibration\), which is consistent with the design goal of this pilot: validating end\-to\-end execution and computational scalability before full downstream\-task optimization\.

### 7\.7Asymptotic Scaling \(Analytical\)

Figure[5](https://arxiv.org/html/2608.26149#S7.F5)plots the*analytical*cost functions of dense attention \(∝n2\\propto n^\{2\}\) and sparse relational attention \(∝n⋅k\\propto n\\cdot k, fixedkk\) established in Proposition[1](https://arxiv.org/html/2608.26149#Thmproposition1)\. These are mathematical functions, not measured runtimes; wall\-clock and memory measurements on Synthea are reported in Section[7\.4](https://arxiv.org/html/2608.26149#S7.SS4)\. We do not include a separate wall\-clock scaling figure in this version\.

nn\(entities, a\.u\.\)relative cost \(a\.u\.\)O​\(n2\)O\(n^\{2\}\)\(dense\)O​\(n⋅k\)O\(n\\cdot k\)\(sparse, fixedkk\)Figure 5:Analytical cost of dense versus sparse relational attention from Proposition[1](https://arxiv.org/html/2608.26149#Thmproposition1)\. Axes are in arbitrary units; this figure depicts the asymptotic functions, not measured performance\.
### 7\.8Progressive Validation Plan

The following plan summarizes completed validation work and the remaining phase required for full clinical confirmation\.

##### Phase 1: Validation on Existing Benchmarks \(Months 1–6\) —largely complete\.

This phase has been executed for open synthetic data\. The full Synthea run \(22,913 patients, 5 seeds×\\times50 epochs, SNOMED CT coding\) is complete; results are reported in Tables[3](https://arxiv.org/html/2608.26149#S7.T3)and[4](https://arxiv.org/html/2608.26149#S7.T4)\. The MIMIC\-IV demo scalability benchmark \(5 seeds, 5 epochs, QUARTER/HALF/FULL subsets\) is complete; results are reported in Table[5](https://arxiv.org/html/2608.26149#S7.T5)\. The module ablation study isolating the marginal contribution of each of the four modules \(M1–M4\) is complete\. The one remaining item in this phase is clinical validation on the*full*MIMIC\-IV dataset \(26 tables,\>\>15M measurements\), which requires PhysioNet credential approval and is scheduled as the immediate next action\.

##### Phase 2: MIMIC\-IV Clinical Validation \+ MT\-5D\-Bench \(Months 7–12\) —pending\.

This phase has two concurrent tracks\. The first track completes the clinical validation deferred from Phase 1: once PhysioNet credentials are approved, the full MIMIC\-IV dataset will be used to validate Modules 2 \(temporal\) and 4 \(causal discovery\) on real dense irregular physiological time series — the setting where synthetic data is least representative\. The second track addresses the standardized benchmark gap identified in Section[2](https://arxiv.org/html/2608.26149#S2): compilation of 10 multi\-table datasets from diverse domains \(healthcare, finance, e\-commerce, IoT, scientific data\), definition of 20 standardized tasks covering all five dimensions, and open\-source release with a public leaderboard\. The benchmark design will ensure that no existing method can achieve high scores across all tasks without addressing all five dimensions\.

##### Phase 3: Industrial Validation \(Year 2\)\.

The third phase transitions from academic benchmarks to production environments through partnerships with healthcare institutions, financial organizations, and retail companies\. Real\-world case studies will measure return on investment, analyst productivity gains, and the number of actionable cross\-table insights discovered compared to existing pipelines\. Scalability testing on production databases exceeding10810^\{8\}records will validate the computational efficiency claims derived from the sparse relational attention complexity analysis\.

### 7\.9Current Status and Remaining Phase

At the date of this manuscript revision, the empirical program has produced two completed result blocks and one remaining validation block\.

##### Completed block A \(synthetic full run\)\.

The end\-to\-end Synthea experiment is complete \(22,913 patients; 1,320,007 encounters; 821,371 condition rows; 5 seeds×\\times50 epochs\)\. Main and ablation results are reported in Tables[3](https://arxiv.org/html/2608.26149#S7.T3)and[4](https://arxiv.org/html/2608.26149#S7.T4)\.

##### Completed block B \(scalability pilot on existing benchmark\)\.

The MIMIC\-IV demo benchmark pilot is complete \(Table[5](https://arxiv.org/html/2608.26149#S7.T5)\), confirming computational scaling behavior and resource use across QUARTER/HALF/FULL subsets\.

##### Remaining validation block \(last empirical phase before industrial deployment\)\.

The last remaining empirical phase is full\-clinical validation on complete MIMIC\-IV \(pending PhysioNet credentials\)\. Its objective is to confirm, on real ICU trajectories, the temporal and relational\-discovery behavior of Modules 2 and 4 under dense irregular measurements\. This block is the immediate next step before large\-scale industrial rollout\.

### 7\.10Phase 2 Progress Achieved Without PhysioNet Access

Without PhysioNet credentials, we completed the maximum feasible Phase 2 actions that do not require restricted clinical data:

- •Open\-data full\-run completion:end\-to\-end Synthea training/evaluation with 5 seeds and 50 epochs, including ablations and baseline comparison \(Tables[3](https://arxiv.org/html/2608.26149#S7.T3)–[4](https://arxiv.org/html/2608.26149#S7.T4)\)\.
- •Scalability pre\-validation on existing benchmark:MIMIC\-IV demo pilot for throughput/compression/runtime/memory across QUARTER/HALF/FULL subsets \(Table[5](https://arxiv.org/html/2608.26149#S7.T5)\)\.
- •Reproducibility package consolidation:executable training harness, fixed metrics, sparse\-attention equivalence validation, and manuscript\-aligned reporting pipeline \(results\.texgeneration\)\.
- •Phase 2 protocol lock\-in:finalized metric definitions, patient\-level split policy, and ablation protocol so that full MIMIC\-IV execution can start immediately once credential access is granted\.

Therefore, the only blocked Phase 2 component is*full\-clinical execution*on restricted MIMIC\-IV tables\. All non\-restricted methodological, experimental, and reporting prerequisites are complete\.

## 8Implications and Perspectives

### 8\.1Toward a Formal Theory of Learning on Temporal Hypergraphs

A fundamental limitation of the current state of the art, identified in Section[2](https://arxiv.org/html/2608.26149#S2), is the absence of a unified theoretical framework for learning on relational data that simultaneously exhibit temporal dynamics and high\-cardinality attributes\. While PAC\-learning theory provides well\-established guarantees for independent and identically distributed data\[[6](https://arxiv.org/html/2608.26149#bib.bib6)\], and while recent work on graph neural networks has begun to characterize their expressive power in terms of the Weisfeiler–Leman hierarchy\[[4](https://arxiv.org/html/2608.26149#bib.bib4)\], no existing result addresses the convergence properties of learning algorithms operating on temporal hypergraphs with heterogeneous node types\.

Our framework motivates the development of such a theory along three axes\. First, extending PAC\-Bayes bounds to the hypergraph setting would provide finite\-sample generalization guarantees for the RHT architecture, accounting for the dependency structure induced by shared hyperedges\. Second, a formal complexity analysis of the sparse relational attention mechanism \(Module 3\) is needed to characterize the trade\-off between computational savings and information loss as a function of graph sparsity\. Third, the differentiable relational discovery process \(Module 4\) raises fundamental questions in optimization theory, as the joint optimization over graph structure and model parameters defines a bilevel optimization problem whose convergence properties remain to be established\.

From a validation standpoint, these theoretical contributions could be assessed through controlled synthetic experiments on random temporal hypergraph models with known ground\-truth properties, allowing direct measurement of the gap between theoretical bounds and empirical performance\. Comparison with existing generalization bounds for GNNs\[[7](https://arxiv.org/html/2608.26149#bib.bib7)\]would provide a concrete baseline for evaluating the tightness of the new guarantees\.

### 8\.2Establishing Relational Data Science as an Interdisciplinary Research Program

Our analysis of the literature reveals that multi\-table learning is currently fragmented across several communities—databases, statistical relational learning, graph neural networks, and temporal modeling—each addressing a subset of the five dimensions in isolation\. As noted in Section[2](https://arxiv.org/html/2608.26149#S2), this fragmentation has resulted in a lack of standardized benchmarks and evaluation protocols, making objective comparison of approaches difficult\.

We argue that addressing this fragmentation requires the establishment of a structured interdisciplinary research program, which we termRelational Data Science\. Unlike existing efforts that combine database theory with machine learning in an ad hoc fashion, this program would be organized around the formal characterization of multidimensional relational complexity and the design of algorithms with provable properties across all five dimensions simultaneously\. The creation of the MT\-5D\-Bench benchmark \(Section[7\.8](https://arxiv.org/html/2608.26149#S7.SS8)\) constitutes a first concrete step in this direction, providing a standardized evaluation infrastructure analogous to the role played by ImageNet in computer vision or GLUE in natural language processing\.

A credible validation of this research direction would involve measuring the adoption and impact of the benchmark within the community: number of participating teams, diversity of proposed methods, and whether the leaderboard reveals systematic performance gaps on specific dimensions that motivate targeted research\. Additionally, cross\-domain transfer experiments—training on healthcare relational data and evaluating on financial or retail datasets—would test whether the proposed formalism genuinely captures domain\-invariant relational structure, as hypothesized in Section[3](https://arxiv.org/html/2608.26149#S3)\.

### 8\.3Industrial Implications and Operational Transformation

From an industrial perspective, the limitations identified in our comparative analysis \(Table[1](https://arxiv.org/html/2608.26149#S2.T1)\) have direct practical consequences\. Current approaches based on table joining followed by standard machine learning \(scoring 11/50 in our multidimensional evaluation\) force organizations into costly and lossy data preparation pipelines, while more sophisticated methods like TGN \(30/50\) require significant expertise and fail to address high cardinality, a pervasive challenge in domains such as healthcare \(ICD\-10 codes\), e\-commerce \(product catalogs\), and finance \(transaction codes\)\.

Our framework addresses this gap by proposing a shift from siloed, table\-level analyses to holistic relational analytics that preserve the semantic richness of inter\-table relationships\. This shift has implications beyond mere technical improvement: it enables new analytical workflows where cross\-table patterns—such as the interaction between prescription history and laboratory trends mediated by diagnostic codes—can be discovered and exploited without manual feature engineering\. The practical feasibility of this transformation would be validated through the industrial partnerships planned in Phase 3 of the validation roadmap \(Section[7\.8](https://arxiv.org/html/2608.26149#S7.SS8)\), with measured indicators including analyst productivity gains, reduction in data preparation time, and the number of actionable cross\-table insights discovered compared to traditional pipelines\.

Furthermore, the modularity of the RHT architecture makes it compatible with emerging data mesh architectures, where data ownership is decentralized across domain teams\. The federated learning component \(Step 7\) directly addresses the challenge of preserving relational coherence across organizational boundaries, a problem for which current data mesh frameworks offer no integrated solution\.

### 8\.4Evolution of Analytical Tools and Platforms

Current business intelligence and data science platforms are fundamentally designed around the single\-table paradigm\. As documented in Section[2](https://arxiv.org/html/2608.26149#S2), even modern platforms that support graph databases or knowledge graphs lack native support for the joint modeling of relational structure, temporal dynamics, and high cardinality that our framework addresses\.

The RHT architecture suggests a new generation of analytical tools built around three capabilities that are absent from the current landscape\. First, relational exploration interfaces that allow analysts to navigate the hypergraph representation interactively, examining how entities are connected across tables and how these connections evolve over time\. Second, relational AutoML systems that automate the eight\-step methodology, from meta\-profiling through deployment, adapting the pipeline configuration to the dimensional profile of the data\. Third, natural language interfaces that translate relational queries \(e\.g\.,“Which patients with rare diagnoses had abnormal lab trends preceding readmission?”\) into operations on the PentE embedding space, bridging the gap between domain expertise and technical implementation\.

The scientific contribution underlying these tools lies in the development of efficient query processing algorithms over the PentE space\. In particular, approximate nearest\-neighbor search in the pentadimensional embedding space requires adaptation of existing indexing structures \(e\.g\., HNSW, IVF\) to account for the heterogeneous metric structure of the five embedding components\. The expected computational complexity and recall guarantees of such adapted structures would be evaluated on the MT\-5D\-Bench datasets, with comparison against exhaustive search and existing multi\-modal retrieval baselines\.

## 9Conclusion

### 9\.1Summary of Contributions

This paper has presented a unified conceptual and methodological framework for multi\-table analysis that addresses five dimensions of complexity simultaneously\. The first contribution is a prospective vision for a unified analytical ecosystem built on relational intelligence, which moves beyond the fragmented approaches documented in our state\-of\-the\-art analysis and proposes an integrated treatment of volume, variables, cardinality, inter\-table relationships, and temporal measurements\. The second contribution is the Relational Hypergraph Transformer architecture, which combines hypergraph representations for n\-ary relationships, pentadimensional embeddings \(PentE\) for unified latent\-space encoding, and adaptive sparse attention mechanisms for computationally efficient cross\-table learning\. The third contribution consists of methodological innovations—relational contrastive learning, dynamic graph rewiring, and relational causal inference—that extend the framework beyond supervised prediction to unsupervised structure discovery and causal reasoning\. The fourth contribution is a comprehensive comparative benchmark that positions our approach against the state of the art across all five dimensions, supported by a formal scoring methodology\. Finally, the fifth contribution is a multidimensional evaluation framework, including novel metrics such as RCR@kkfor rare category assessment and the 5D Integration Score based on harmonic means, which penalizes approaches that neglect any single dimension\.

### 9\.2Decisive Advantages of the Proposed Approach

Three structural advantages distinguish our framework from existing methods\. First, the simultaneous processing of all five dimensions through a single unified architecture avoids the cascading information loss inherent in sequential pipelines—where, for instance, flattening relational structure before addressing temporality destroys relational semantics that are irrecoverable downstream\. The theoretical complexity reduction fromO​\(n2\)O\(n^\{2\}\)toO​\(n⋅k\)O\(n\\cdot k\)achieved by sparse relational attention ensures that this holistic treatment does not incur prohibitive computational costs, a critical requirement for scalability to production\-scale relational databases\. Second, the domain\-agnostic design of the framework, grounded in the formal definition of the relational hypergraph \(Definition 1\) and the PentE embedding space, ensures applicability across healthcare, finance, retail, IoT, and scientific domains without requiring domain\-specific architectural modifications\. Third, the preservation of relational structure throughout the analytical pipeline—from hypergraph construction through deployment—provides a natural basis for prediction explainability, as the attention weights and hyperedge activations can be traced back to specific inter\-table relationships\.

### 9\.3Limitations and Directions for Future Research

Several limitations of the current work must be acknowledged, each pointing toward specific research directions\.

The first limitation concerns implementation complexity\. The RHT architecture involves multiple interacting components whose joint behavior may exhibit emergent difficulties not apparent from the analysis of individual modules\. In particular, the interaction between differentiable hypergraph pruning \(Module 1\) and the downstream contrastive learning objective \(Step 4\) creates a training dynamics that may be sensitive to initialization and hyperparameter choices\. Future work should investigate the stability and convergence properties of this joint optimization, drawing on recent advances in bilevel optimization theory, and develop principled initialization strategies informed by the meta\-profiling stage \(Step 1\)\.

The second limitation relates to computational cost\. While sparse relational attention reduces asymptotic complexity, the initial hypergraph construction and PentE embedding computation remain resource\-intensive for very large databases\. Current temporal graph networks such as TGN\[[7](https://arxiv.org/html/2608.26149#bib.bib7)\]face similar scaling challenges, typically demonstrated on graphs with at most10610^\{6\}edges\. Extending our framework to databases with10810^\{8\}or more records will require algorithmic innovations in approximate hypergraph construction and distributed PentE computation\. A promising direction is the adaptation of locality\-sensitive hashing techniques to the hypergraph setting, which could reduce construction complexity fromO​\(\|𝒱\|2\)O\(\|\\mathcal\{V\}\|^\{2\}\)to near\-linear time while preserving the most informative hyperedges\.

The third limitation is the dependency on relational metadata quality\. The framework assumes that meaningful relational structure—explicit or latent—exists in the data and can be discovered by Module 4\. In degenerate cases where tables share no semantic or structural overlap, the hypergraph representation reduces to a set of disconnected components, and the benefits of cross\-table learning diminish\. Characterizing the conditions under which multi\-table analysis provably outperforms independent per\-table analysis is an open theoretical question that connects to the broader literature on transfer learning and multi\-task learning bounds\.

The fourth limitation is that the empirical evaluation \(Tables[3](https://arxiv.org/html/2608.26149#S7.T3)–[4](https://arxiv.org/html/2608.26149#S7.T4)\) is conducted on open synthetic data \(Synthea, SNOMED CT coding, 22,913 patients\) rather than real clinical EHR data\. While the results are fully measured and reproducible, synthetic patients do not exhibit the temporal complexity, comorbidity patterns, or measurement noise characteristic of real ICU data\. In particular, Modules 2 \(multi\-scale temporal embeddings\) and 4 \(relational causal discovery\) are exercised only shallowly on encounter timestamps; their validation on the dense, irregular physiological time series of MIMIC\-IV \(laboratory trends, vital signs at sub\-hourly resolution\) remains the essential remaining step\. Clinical validation on full MIMIC\-IV, pending PhysioNet credential approval, will determine whether the semantic coherence advantage demonstrated on Synthea transfers to a real clinical population\.

Looking beyond these immediate limitations, three longer\-term research directions merit investigation\. Extension to multimodal knowledge graphs—integrating textual clinical notes, medical images, and tabular records into a single hypergraph—would address the increasingly multimodal nature of real\-world data, though it requires fundamental advances in cross\-modal alignment within the PentE space\. Integration with large language models for natural relational query generation could democratize access to complex multi\-table analyses, but raises challenges in grounding language model outputs in the formal hypergraph structure\. Finally, establishing PAC\-style learning guarantees for temporal hypergraph models would provide the theoretical foundation needed for safety\-critical applications in healthcare and finance\.

### 9\.4Research Roadmap

The realization of this research program is organized in three phases aligned with the validation plan \(Section[7\.8](https://arxiv.org/html/2608.26149#S7.SS8)\)\. The first phase \(largely complete\) has delivered: \(i\) an open\-source PyTorch implementation of the full RHT pipeline; \(ii\) a Synthea full run with confirmed empirical measurements \(Tables[3](https://arxiv.org/html/2608.26149#S7.T3)–[4](https://arxiv.org/html/2608.26149#S7.T4)\); \(iii\) a MIMIC\-IV demo scalability benchmark \(Table[5](https://arxiv.org/html/2608.26149#S7.T5)\); and \(iv\) a module ablation study\. The one remaining Phase 1 item is clinical validation on full MIMIC\-IV \(pending PhysioNet credentials\)\. During months seven through twelve \(Phase 2\), all non\-restricted tasks are already in place \(protocols, metrics, ablations, reproducible harness\), so execution can proceed immediately to full\-clinical runs once credentials are approved; in parallel, ecosystem integration and scalability work \(deployment optimizations, pre\-trained relational schemas, and relational AutoML\) continues\. The second year targets industrial validation through partnerships with healthcare institutions and financial organizations, producing real\-world case studies with measured return on investment and scalability assessments on production\-scale databases exceeding10810^\{8\}records\.

### 9\.5General Conclusion

Multi\-table analysis represents a critical frontier in artificial intelligence and data science\. The true complexity of modern data does not reside in an isolated dimension, but in the sophisticated interaction of multiple dimensions: volume, variety, cardinality, relationships, and temporality\.

Our 5D methodological framework recognizes this reality and proposes a systemic, holistic, and technically innovative solution\. By simultaneously addressing these five dimensions through a unified architecture, we pave the way toward a new generation of analytical systems capable of extracting the full informational richness of complex relational data\.

We provide an open implementation and an empirical evaluation of the high\-cardinality components of the architecture on synthetic EHR data, together with a proven complexity bound for the core attention operator\. The measured results \(Section[7\.4](https://arxiv.org/html/2608.26149#S7.SS4)\) and ablation \(Section[7\.5](https://arxiv.org/html/2608.26149#S7.SS5)\) characterise the contribution of each module; clinical validation on MIMIC\-IV is the immediate next step\. More fundamentally, this work lays the foundations for an emerging discipline at the intersection of databases, machine learning, and temporal analysis, with implications for academic research, industry, and data science practitioners\. The path toward complete industrialization of these methods requires sustained effort in both theoretical development and empirical validation, but the direction is clear: the future of analytics belongs to approaches that embrace relational complexity rather than circumvent it\.

## Acknowledgments

The authors thank the maintainers of Synthea and the open\-source scientific Python ecosystem for enabling reproducible experimentation\.

## References

- \[1\]Kimball, R\., & Ross, M\. \(2013\)\.The data warehouse toolkit: The definitive guide to dimensional modeling\(3rd ed\.\)\. John Wiley & Sons\.
- \[2\]Singh, A\. P\., & Gordon, G\. J\. \(2008\)\. Relational learning via collective matrix factorization\. InProceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining\(pp\. 650\-658\)\.
- \[3\]Richardson, M\., & Domingos, P\. \(2006\)\. Markov logic networks\.Machine Learning, 62\(1\-2\), 107\-136\.
- \[4\]Kipf, T\. N\., & Welling, M\. \(2017\)\. Semi\-supervised classification with graph convolutional networks\. InInternational Conference on Learning Representations \(ICLR\)\.
- \[5\]Walonoski, J\., Kramer, M\., Nichols, J\., et al\. \(2018\)\. Synthea: An approach, method, and software mechanism for generating synthetic patients and the synthetic electronic health care record\.Journal of the American Medical Informatics Association, 25\(3\), 230–238\.
- \[6\]Hamilton, W\., Ying, Z\., & Leskovec, J\. \(2017\)\. Inductive representation learning on large graphs\. InAdvances in Neural Information Processing Systems\(pp\. 1024\-1034\)\.
- \[7\]Rossi, E\., Chamberlain, B\., Frasca, F\., Eynard, D\., Monti, F\., & Bronstein, M\. \(2020\)\. Temporal graph networks for deep learning on dynamic graphs\.arXiv preprint arXiv:2006\.10637\.
- \[8\]Xu, D\., Ruan, C\., Korpeoglu, E\., Kumar, S\., & Achan, K\. \(2020\)\. Inductive representation learning on temporal graphs\. InInternational Conference on Learning Representations \(ICLR\)\.
- \[9\]Peters, M\. E\., Neumann, M\., Logan IV, R\. L\., Schwartz, R\., Joshi, V\., Singh, S\., & Smith, N\. A\. \(2019\)\. Knowledge enhanced contextual word representations\. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing\.
- \[10\]Kazemi, S\. M\., Goel, R\., Eghbali, S\., Ramanan, J\., Sahota, J\., Thakur, S\., … & Brubaker, M\. A\. \(2019\)\. Time2vec: Learning a vector representation of time\.arXiv preprint arXiv:1907\.05321\.

## Appendix ADetailed Pseudo\-code

### A\.1Hypergraph Construction

Listing 1:Detailed implementation of hypergraph construction1classAdaptiveHypergraphConstructor:

2def\_\_init\_\_\(self,tables,metadata\):

3self\.tables=tables

4self\.metadata=metadata

5self\.relations=\[\]

6self\.hypergraph=None

7

8defdetect\_relations\_multi\_level\(self\):

9"""Detectrelationsatmultiplelevels"""

10

11explicit\_rels=self\.\_detect\_foreign\_keys\(\)

12

13

14implicit\_rels=self\.\_detect\_value\_overlaps\(\)

15

16

17semantic\_rels=self\.\_infer\_semantic\_relations\(\)

18

19

20temporal\_rels=self\.\_detect\_temporal\_patterns\(\)

21

22self\.relations=\{

23’explicit’:explicit\_rels,

24’implicit’:implicit\_rels,

25’semantic’:semantic\_rels,

26’temporal’:temporal\_rels

27\}

28returnself\.relations

29

30defconstruct\_hypergraph\(self\):

31"""Constructthehypergraph"""

32nodes=self\.\_extract\_entities\(\)

33hyperedges=self\.\_create\_nary\_edges\(self\.relations\)

34weights=self\.\_compute\_edge\_weights\(hyperedges\)

35

36self\.hypergraph=HyperGraph\(

37nodes=nodes,

38hyperedges=hyperedges,

39weights=weights

40\)

41returnself\.hypergraph

42

43defdifferentiable\_pruning\(self,threshold=0\.1\):

44"""Differentialgraphpruning"""

45

46importance=self\.\_compute\_edge\_importance\(\)

47

48

49pruned\_edges=\[\]

50foredge,scoreinzip\(self\.hypergraph\.hyperedges,importance\):

51ifscore\>threshold:

52pruned\_edges\.append\(edge\)

53

54self\.hypergraph\.hyperedges=pruned\_edges

55returnself\.hypergraph

### A\.2PentE Embeddings

Listing 2:Computation of pentadimensional embeddings1classPentEEmbedding\(nn\.Module\):

2def\_\_init\_\_\(self,config\):

3super\(\)\.\_\_init\_\_\(\)

4self\.semantic\_encoder=SemanticEncoder\(config\.semantic\_dim\)

5self\.relational\_encoder=RelationalGNN\(config\.relation\_dim\)

6self\.temporal\_encoder=Time2Vec\(config\.temporal\_dim\)

7self\.categorical\_encoder=HierarchicalCatEncoder\(config\.cat\_dim\)

8self\.volume\_normalizer=VolumeNormalizer\(config\.vol\_dim\)

9

10defforward\(self,entity,graph,timestamp,categories\):

11

12sem\_emb=self\.semantic\_encoder\(entity\.attributes\)

13

14

15rel\_emb=self\.relational\_encoder\(entity,graph\)

16

17

18temp\_emb=self\.temporal\_encoder\(timestamp\)

19

20

21cat\_emb=self\.categorical\_encoder\(categories\)

22

23

24vol\_emb=self\.volume\_normalizer\(entity\.volume\_stats\)

25

26

27pente\_emb=torch\.cat\(\[

28sem\_emb,rel\_emb,temp\_emb,cat\_emb,vol\_emb

29\],dim=\-1\)

30

31returnpente\_emb

## Appendix BDatasets and Code

Source code, datasets, and pre\-trained models will be made available at:

https://github\.com/edlansiaux/multitable\-5d\-analysis

The MT\-5D\-Bench benchmarks will be hosted at:

https://mt5d\-benchmark\.org