Drift-Aware Temporal Graph Rewiring (DATGR) for Adaptive Semantic Modeling in Biomedical Text

# Drift-Aware Temporal Graph Rewiring (DATGR) for Adaptive Semantic Modeling in Biomedical Text
Source: [https://arxiv.org/html/2607.08490](https://arxiv.org/html/2607.08490)
Bharathwaj Vijayakumar1and Sahana K\. Varadaraju21Bharathwaj Vijayakumar is with Institutional Data and Analytics, Rowan University, NJ, USA\. \(e\-mail: vijayakumar@rowan\.edu\)2Sahana K\. Varadaraju is with Information Resources & Technology, Rowan University, NJ, USA\. \(e\-mail: varadaraju@rowan\.edu\)© 2026 IEEE\. Personal use of this material is permitted\. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works\. DOI: 10\.1109/CAI68641\.2026\.11536228

###### Abstract

Biomedical language evolves rapidly as new discoveries emerge, causing traditional text models to lose semantic fidelity over time\. Static embeddings and co\-occurrence graphs cannot capture such evolution, leading to performance degradation in retrieval and knowledge discovery tasks\. This paper introduces aDrift\-Aware Temporal Graph Rewiring \(DATGR\)framework that models concept evolution by dynamically updating co\-occurrence edges based on estimated semantic drift\. Instead of retraining embeddings for each time slice, DATGR performs lightweight, feedback\-driven rewiring using a logistic update rule applied to edge weights\. Evaluated on the Biomedical Multi\-Relation Corpus \(BIOMRC\), the method achieved a mean Area Under the Receiver Operating Characteristic \(AUROC\) improvement of approximately 0\.066 absolute difference \(0\.6990\.699vs\.0\.6330\.633\) over a static baseline\. Area Under the Precision\-Recall Curve \(AUPRC\) remained comparable \(0\.7380\.738vs\.0\.7440\.744\), showing that drift\-aware adaptation enhances link\-prediction recall without a loss in precision\. These results demonstrate that edge\-level adaptation effectively captures temporal semantic change in evolving biomedical text while remaining computationally efficient and interpretable\.

## IINTRODUCTION

The exponential growth of biomedical publications has transformed scientific text into a continuously evolving knowledge ecosystem\. Thousands of new abstracts appear weekly in PubMed and related repositories, introducing novel concepts, drug\-compound interactions, and terminologies that shift the semantic structure of the domain\. Capturing such concept evolution is critical for applications in literature mining, knowledge graph curation, and retrieval\-augmented generation\. However, most graph\-based or embedding\-based text models still assume static relationships between words and entities, which quickly become outdated as meaning drifts over time\.

Traditional co\-occurrence graphs\[[1](https://arxiv.org/html/2607.08490#bib.bib1)\]and word embeddings such as word2vec\[[2](https://arxiv.org/html/2607.08490#bib.bib2)\]or GloVe\[[3](https://arxiv.org/html/2607.08490#bib.bib3)\]treat semantics as stationary\. In reality, biomedical terms like “immune checkpoint,” “mRNA vaccine,” or “viral vector” evolve as research advances, resulting in measurable semantic drift\[[4](https://arxiv.org/html/2607.08490#bib.bib4),[5](https://arxiv.org/html/2607.08490#bib.bib5),[7](https://arxiv.org/html/2607.08490#bib.bib7)\]\. Static models therefore misrepresent emerging associations and reduce retrieval accuracy when used for downstream analysis\.

To mitigate this, several temporal embedding frameworks\[[6](https://arxiv.org/html/2607.08490#bib.bib6),[7](https://arxiv.org/html/2607.08490#bib.bib7)\]and dynamic graph neural networks \(DGNNs\), and surveys thereof\[[8](https://arxiv.org/html/2607.08490#bib.bib8),[10](https://arxiv.org/html/2607.08490#bib.bib10)\], have been proposed\. While effective, many methods require retraining or sequential fine\-tuning for each time slice; these approaches incur significant computational cost and limiting interpretability\. Furthermore, most focus on node\-level drift, how individual term vectors change, rather than on edge\-level drift, which captures the evolving co\-occurrence structure among terms\. Instead of retraining embeddings, DATGR rewires the word\-co\-occurrence graph using a drift\-weighted logistic update rule, enabling continuous adaptation without full re\-embedding\. Evaluated on theBIOMRCcorpus, the proposed model achieved an absolute AUROC improvement of approximately 0\.066 \(0\.6990\.699vs\.0\.6330\.633\) over a static baseline, while maintaining comparable AUPRC performance \(0\.7380\.738vs\.0\.7440\.744\)\. These results demonstrate that lightweight, feedback\-driven edge updates can effectively capture semantic evolution while preserving precision\.

The key contributions of this work are as follows:

- •A drift\-aware rewiring mechanism for temporal co\-occurrence graphs that models edge evolution efficiently without retraining embeddings \(DATGR\);
- •An empirical demonstration on real biomedical text, showing improved link\-prediction AUROC and stable precision\-recall behavior across temporal windows;
- •A scalable foundation for adaptive semantic modeling applicable to evolving domains such as biomedical knowledge tracking, retrieval augmentation, and ontology maintenance\.

## IIRELATED WORK

Research on modeling temporal change in language has evolved along three major directions: static co\-occurrence graphs, temporal word embeddings, and dynamic graph learning\.

Early approaches based on word co\-occurrence statistics treated term relations as static\[[1](https://arxiv.org/html/2607.08490#bib.bib1)\]\. While such representations capture word associations effectively at a single point in time, they ignore how meaning shifts as domains evolve\. Distributional embeddings such as word2vec\[[2](https://arxiv.org/html/2607.08490#bib.bib2)\]and GloVe\[[3](https://arxiv.org/html/2607.08490#bib.bib3)\]improved semantic representation but remained time\-invariant\. Recent work on temporal embeddings for biomedical literature has explored how contextual meaning changes across time\[[11](https://arxiv.org/html/2607.08490#bib.bib11)\]\. To address temporal variability, researchers proposed diachronic or dynamic embeddings that explicitly model meaning change across time\. Hamiltonet al\.\[[4](https://arxiv.org/html/2607.08490#bib.bib4)\]analyzed word embedding trajectories to uncover statistical laws of semantic change, while Tahmasebiet al\.\[[5](https://arxiv.org/html/2607.08490#bib.bib5)\]surveyed computational methods for detecting lexical semantic change\. Bamler and Mandt\[[6](https://arxiv.org/html/2607.08490#bib.bib6)\]introduced probabilistic dynamic embeddings, modeling word transitions as latent trajectories through time\. Nicholsonet al\.\[[7](https://arxiv.org/html/2607.08490#bib.bib7)\]analyzed large\-scale biomedical corpora and demonstrated domain\-specific semantic drift in biomedical scientific language\.

Dynamic graph learning addresses temporal networks directly\. Recent surveys of DGNNs and temporal knowledge graphs \(e\.g\.,\[[8](https://arxiv.org/html/2607.08490#bib.bib8)\]\) summarize architectures that update node representations as edges arrive\[[8](https://arxiv.org/html/2607.08490#bib.bib8)\]\. However, many such methods emphasize node embeddings rather than direct edge updates reflecting co\-occurrence strength\. The proposed DATGR differs by explicitly modeling edge drift via a lightweight logistic rewiring rule, achieving temporal sensitivity while remaining interpretable and efficient\.

## IIIMETHODOLOGY

The proposed DATGR framework models the evolution of biomedical semantics through incremental, edge\-level graph adaptation rather than full re\-embedding\. It comprises four modules: \(1\) temporal segmentation of the corpus, \(2\) drift estimation using sentence embeddings, \(3\) drift\-weighted graph rewiring, and \(4\) link prediction with node embeddings\. Fig\.[1](https://arxiv.org/html/2607.08490#S3.F1)illustrates this pipeline\.

![Refer to caption](https://arxiv.org/html/2607.08490v1/system_diagram.png)Figure 1:Overview of the DATGR \(drift\-aware temporal graph rewiring\) framework, showing the stages of temporal segmentation, drift estimation, graph rewiring, and link\-prediction evaluation\.### III\-ATemporal Segmentation and Graph Construction

The biomedical corpus is divided intoTTchronological windows\{W1,W2,…,WT\}\\\{W\_\{1\},W\_\{2\},\\dots,W\_\{T\}\\\}, each representing a distinct semantic period\. For every window, tokenized abstracts are converted into an undirected weighted graphGt=\(Vt,Et,Wt\)G\_\{t\}=\(V\_\{t\},E\_\{t\},W\_\{t\}\)whereVtV\_\{t\}represents high\-frequency biomedical terms andEtE\_\{t\}encodes co\-occurrence within a context window of sizenn\. Each edge weight is normalized by the maximum observed frequency:

Wt​\(i,j\)=Count​\(i,j\)maxa,b⁡Count​\(a,b\)\.W\_\{t\}\(i,j\)=\\frac\{\\text\{Count\}\(i,j\)\}\{\\max\_\{a,b\}\\text\{Count\}\(a,b\)\}\.\(1\)This normalization allows comparability of edge strengths across windows with differing corpus sizes\.

### III\-BDrift\-Aware Rewiring Rule

DATGR introduces a logistic edge update rule that adjusts edge weights between two terms\(i,j\)\(i,j\)using their prior weight, current co\-occurrence strength, and local drift\.

We denote byst​\(i,j\)=Wt​\(i,j\)s\_\{t\}\(i,j\)=W\_\{t\}\(i,j\)the normalized co\-occurrence strength in the current windowtt\(cf\. Eq\. \([1](https://arxiv.org/html/2607.08490#S3.E1)\)\), and define the pairwise change

Δi​j=st​\(i,j\)−st−1​\(i,j\)\.\\Delta\_\{ij\}\\;=\\;s\_\{t\}\(i,j\)\\;\-\\;s\_\{t\-1\}\(i,j\)\.\(2\)Given the drift magnitudesDt​\(⋅\)D\_\{t\}\(\\cdot\), the edge update fromt−1t\{\-\}1tottis

W^t​\(i,j\)=\\displaystyle\\hat\{W\}\_\{t\}\(i,j\)\\;=\(1−η\)​Wt−1​\(i,j\)\\displaystyle\(1\-\\eta\)\\,W\_\{t\-1\}\(i,j\)\(3\)\+ησ\(b0\+b1st\(i,j\)\+b2Δi​j\\displaystyle\+\\;\\eta\\,\\sigma\\\!\\Big\(b\_\{0\}\\;\+\\;b\_\{1\}\\,s\_\{t\}\(i,j\)\\;\+\\;b\_\{2\}\\,\\Delta\_\{ij\}\+b3\[Dt\(i\)\+Dt\(j\)\]\),\\displaystyle\\qquad\\qquad\\qquad\\;\\;\+\\;b\_\{3\}\\,\[D\_\{t\}\(i\)\+D\_\{t\}\(j\)\]\\Big\),whereσ​\(x\)=1/\(1\+e−x\)\\sigma\(x\)=1/\(1\+e^\{\-x\}\)is the logistic transform andη∈\[0,1\]\\eta\\in\[0,1\]is the adaptation rate\. The first term preserves inertia through a convex combination ofWt−1W\_\{t\-1\}, while the second term injects innovation using current evidence \(sts\_\{t\}\), its change \(Δi​j\\Delta\_\{ij\}\), and the local drift signal \(DtD\_\{t\}\)\. BecauseWt−1∈\[0,1\]W\_\{t\-1\}\\in\[0,1\]andσ​\(⋅\)∈\(0,1\)\\sigma\(\\cdot\)\\in\(0,1\), the convex combination guaranteesW^t​\(i,j\)∈\[0,1\]\\hat\{W\}\_\{t\}\(i,j\)\\in\[0,1\]\.

#### Practical notes

We foundη∈\[0\.1,0\.2\]\\eta\\\!\\in\\\!\[0\.1,0\.2\]to balance stability and responsiveness\. Coefficients\(b0,b1,b2,b3\)\(b\_\{0\},b\_\{1\},b\_\{2\},b\_\{3\}\)can be tuned on a validation slice; settingb2=0b\_\{2\}\{=\}0reduces the rule to a drift and strength\-weighted update when change terms are noisy\. A small threshold onDt​\(⋅\)D\_\{t\}\(\\cdot\)may be applied to ignore negligible drift\.

### III\-CTop\-k Sparsification and Stability

After rewiring, each node retains only its top\-k neighbors ranked byW^t​\(i,j\)\\hat\{W\}\_\{t\}\(i,j\)\. This enforces sparsity and mitigates noise propagation due to spurious term co\-occurrences\. Empirically, k=5 provided the best trade\-off between stability and adaptability, yielding consistent graph density across time\.

### III\-DNode Embedding and Link Prediction

Updated graphsG^t\\hat\{G\}\_\{t\}are embedded via Node2Vec to derive latent representationsZuZ\_\{u\}for every nodeuu\. Each candidate edge\(u,v\)\(u,v\)is represented as the element\-wise product:

zu​v=Zu⊙Zv\.z\_\{uv\}=Z\_\{u\}\\odot Z\_\{v\}\.\(4\)A logistic regression classifier is trained to predict whether\(u,v\)\(u,v\)will appear inGt\+1G\_\{t\+1\}, effectively testing the model’s ability to anticipate new semantic associations\.

### III\-EComputational Complexity

Let\|Et\|\|E\_\{t\}\|denote the number of edges per window\. DATGR’s update complexity isO​\(\|Et\|\)O\(\|E\_\{t\}\|\)since each edge weight is modified independently, avoiding theO​\(\|V\|2\)O\(\|V\|^\{2\}\)cost of re\-embedding or backpropagation through time\. This results in substantially lower computational cost than dynamic GNN approaches while preserving a comparable temporal sensitivity\.

### III\-FAlgorithm Summary

Algorithm 1Drift\-Aware Temporal Graph Rewiring \(DATGR\)\.1:Split corpus into windows

\{W1,…,WT\}\\\{W\_\{1\},\\dots,W\_\{T\}\\\}
2:for

t=2t=2to

TTdo

3:Construct co\-occurrence graph

GtG\_\{t\}
4:Compute term drift

Dt​\(w\)D\_\{t\}\(w\)
5:foreach edge

\(i,j\)∈Gt−1\(i,j\)\\in G\_\{t\-1\}do

6:Update weight

W^t​\(i,j\)\\hat\{W\}\_\{t\}\(i,j\)using Eq\. \([3](https://arxiv.org/html/2607.08490#S3.E3)\)

7:endfor

8:Retain top\-k edges per node

9:endfor

10:Train Node2Vec on

G^t\\hat\{G\}\_\{t\}; predict edges in

Gt\+1G\_\{t\+1\}

## IVEXPERIMENTS AND RESULTS

### IV\-ADataset and Temporal Segmentation

Experiments used theBIOMRCdataset\[[14](https://arxiv.org/html/2607.08490#bib.bib14)\], divided intoT=4T=4windows of approximately 1000 abstracts each, forming pseudo\-chronological snapshots\.

### IV\-BGraph Construction

For each window, the 400 most frequent biomedical terms formedVtV\_\{t\}, yielding∼\\sim400 nodes and 4k\-5k edges\. Co\-occurrence weights followed Eq\. \([1](https://arxiv.org/html/2607.08490#S3.E1)\)\.

### IV\-CSemantic Drift Estimation

Embeddings were produced with all\-MiniLM\-L6\-v2\. Drift between windows was computed using cosine distance between consecutive term representations\.

### IV\-DBaselines and Configuration

A static graph baseline \(fixed edges\) was compared to DATGR with identical settings:η=0\.2\\eta=0\.2,\(b0,b1,b2,b3\)=\(0,3,2,1\)\(b\_\{0\},b\_\{1\},b\_\{2\},b\_\{3\}\)=\(0,3,2,1\), top\-k=5; Node2Vec dim=64, walk length=8, 20 walks/node, context window=5, seed=42\.

### IV\-EQuantitative Results

DATGR and the static baseline were compared using link\-prediction performance across consecutive temporal windows\. Performance was measured via AUROC and AUPRC, averaged across three consecutive window transitions \(0→1, 1→2, 2→3\)\. Table[I](https://arxiv.org/html/2607.08490#S4.T1)summarizes the results\.

TABLE I:Static vs\. Drift\-Aware \(DATGR\) Link Prediction Performance\.![Refer to caption](https://arxiv.org/html/2607.08490v1/fig_auroc.png)Figure 2:AUROC across temporal windows\. DATGR consistently outperforms the static baseline, achieving gains of approximately 0\.066 per window\.![Refer to caption](https://arxiv.org/html/2607.08490v1/fig_auprc.png)Figure 3:AUPRC across temporal windows\. DATGR remains comparable to the static model, indicating improved recall without reducing precision\.![Refer to caption](https://arxiv.org/html/2607.08490v1/fig_edits.png)Figure 4:Edge edits per temporal window after DATGR rewiring, demonstrating stable yet adaptive structural updates\.
### IV\-FAnalysis of Figures and Results

Fig\.[2](https://arxiv.org/html/2607.08490#S4.F2)shows that DATGR achieves a consistent AUROC improvement across all windows, achieving an average absolute gain of approximately 0\.066 compared to the static model\. This indicates that our proposed method improves the model’s ability to discriminate between emerging and obsolete semantic relations\. The DATGR curve remains stable, reflecting robust generalization across time slices rather than overfitting to a single window\.

In Fig\.[3](https://arxiv.org/html/2607.08490#S4.F3), both static and drift\-aware models achieve similar AUPRC levels \(∼0\.74\\sim 0\.74\)\. DATGR slightly surpasses the baseline in later windows, while remaining comparable overall\. This stability demonstrates that DATGR preserves precision even as it adapts to new connections, suggesting balanced recall expansion\.

Fig\.[4](https://arxiv.org/html/2607.08490#S4.F4)illustrates that DATGR modifies a consistent, moderate number of edges in each temporal transition, confirming that adaptation occurs without destabilizing the overall structure\. This “steady rewiring” pattern aligns with the theoretical design: a small but meaningful number of edge updates per window efficiently encode semantic evolution while retaining the corpus’s structural coherence\.

Overall, these findings confirm that DATGR effectively balances precision and adaptability\. Its gains in AUROC and stable AUPRC suggest a robust capacity to capture semantic drift while avoiding overfitting or excessive rewiring, key traits for scalable biomedical semantic modeling\.

## VDiscussion and Future Work

### V\-ANovelty and Theoretical Contribution

Most existing frameworks for temporal language modeling focus on either \(1\) retraining embeddings for each time slice\[[4](https://arxiv.org/html/2607.08490#bib.bib4),[6](https://arxiv.org/html/2607.08490#bib.bib6)\], or \(2\) employing recurrent or attention\-based dynamic graph neural networks to propagate node states through time\[[7](https://arxiv.org/html/2607.08490#bib.bib7),[8](https://arxiv.org/html/2607.08490#bib.bib8)\]\. While these approaches successfully capture temporal variation, they typically incur higher computational cost, reduced interpretability, and limited ability to isolate the specific contribution of drift to the evolving network structure\.

In contrast, DATGR introduces an edge\-centric paradigm for temporal adaptation, conceptually related to prior edge\-rewiring approaches such as DiffWire\[[9](https://arxiv.org/html/2607.08490#bib.bib9)\], but extended to incorporate temporal drift signals\. Rather than recalculating entire node embeddings, DATGR explicitly updates the edge weights in the co\-occurrence graph as a closed\-form function of estimated drift\. This shift from node\-level to edge\-level adaptation forms the central novelty of the framework, enabling the model to reflect emerging or diminishing associations between biomedical terms without relearning the entire semantic space\. As a result, semantic evolution is encoded directly in graph structure rather than distributed across latent embedding dimensions, making individual changes traceable and auditable\.

Another notable aspect lies in the logistic update rule, which balances historical persistence with drift\-driven innovation\. Whereas most temporal embedding models rely on additive or linear updates that treat all changes as equally significant, the logistic function in DATGR introduces a saturating non\-linearity that dampens noise while emphasizing strong, consistent shifts in co\-occurrence patterns\. This formulation provides an intuitive regularization effect: large, stable drifts dominate small, transient fluctuations\.

### V\-BEmpirical and Practical Significance

The experimental results indicate that DATGR effectively captures semantic drift while maintaining precision\. Across all temporal windows, DATGR improves AUROC by approximately 0\.066 compared with the static baseline, demonstrating stronger discriminative capability in identifying emerging relations\. AUPRC values remain comparable to the baseline and slightly surpass it in later windows, indicating that recall increases without degrading precision\.

This balance is particularly valuable for biomedical systems where false associations can propagate misinformation\. The ability to achieve AUROC gains while preserving AUPRC parity suggests that DATGR’s rewiring introduces meaningful structural corrections rather than indiscriminate connectivity expansion\. The edge\-edit patterns further reveal steady, controlled adaptation, where each temporal transition modifies only a modest subset of edges, sufficient to capture semantic novelty while preserving network coherence\.

From an operational standpoint, this represents an attractive trade\-off for large\-scale biomedical knowledge platforms such as PubMed, bioRxiv, or ontology\-curation systems\. DATGR can be inserted as a modular update layer between corpus ingestion and embedding generation, requiring no retraining of downstream encoders\. Because the algorithm scales linearly with the number of edgesO​\(\|E\|\)O\(\|E\|\)and avoids backpropagation, it remains computationally feasible for large vocabularies\. Recent benchmarks such as Know2BIO\[[12](https://arxiv.org/html/2607.08490#bib.bib12)\]further highlight the importance of evaluating evolving biomedical knowledge graphs under temporal dynamics – reinforcing the practical value of drift\-aware approaches like DATGR for real\-world corpora\.

Its modular design also makes it well\-suited for retrieval\-augmented generation \(RAG\) pipelines\[[13](https://arxiv.org/html/2607.08490#bib.bib13)\], where maintaining semantically coherent retrieval vectors over time is crucial for factual accuracy\. DATGR can function as a drift\-regularization layer, helping ensure that evolving biomedical concepts remain temporally consistent without retraining large encoder models\.

### V\-CComparison with Previous Paradigms

Table[II](https://arxiv.org/html/2607.08490#S5.T2)conceptually contrasts DATGR with representative temporal modeling paradigms\.

TABLE II:Comparison Between DATGR and Prior Temporal Models\.Abbreviations:Dyn\.=Dynamic; Embed\.=Embedding; Retrain\.=Retraining; Interp\.=Interpretability; Mod\.=Moderate\.

As shown, DATGR provides an interpretable, retraining\-free mechanism that captures temporal sensitivity at the relational level\. It complements rather than replaces neural or embedding\-based paradigms, serving as a stable substrate upon which expressive temporal GNNs or continual learning models can build\.

### V\-DInterpretability and Theoretical Perspective

Beyond quantitative metrics, DATGR’s rewiring dynamics yield a transparent lens into biomedical language evolution\. Each edge weight explicitly reflects the co\-occurrence strength between two terms and evolves according to measurable semantic drift\. Visualizing these rewiring trajectories reveals how associations such as “mRNA vaccine” intensified after 2020 as the term displaced the earlier, broader “RNA vaccine” usage, or how “viral vector” co\-occurrences waned as mRNA\-based delivery methods rose to prominence\. This explicit traceability, linking edge changes directly to drift magnitudes, offers interpretability rarely achievable with latent temporal embeddings or neural recurrent models\.

### V\-EFuture Research Directions

Future work can extend DATGR along several methodological and application axes:

1. 1\.Learnable Parameterization:Introduce adaptive or learnable coefficients\(b0,b1,b2,b3\)\(b\_\{0\},b\_\{1\},b\_\{2\},b\_\{3\}\)so the update rule can calibrate itself dynamically to different domains or drift intensities, effectively creating a “drift\-aware optimizer\.”
2. 2\.Multimodal Drift Integration:Combine textual drift with external modalities such as citation networks or MeSH ontologies to distinguish linguistic change from topic diffusion\.
3. 3\.Hybridization with Neural Temporal Models:Use DATGR as a structural prior for temporal GNNs, where rewired edges guide attention toward semantically active relationships while retaining interpretability\.
4. 4\.Evaluation on True Chronological Corpora:Apply DATGR to genuine temporal datasets \(e\.g\., PubMed 2000–2024, COVID\-19 literature\) to evaluate its ability to track known paradigm shifts in biomedical science, following recent advances in temporal biomedical knowledge graph modeling\[[16](https://arxiv.org/html/2607.08490#bib.bib16)\]\.
5. 5\.Continual RAG Integration:Integrate DATGR within retrieval\-augmented generation pipelines to maintain synchronization between evolving knowledge graphs and static LLM representations\.

### V\-FBroader Implications

DATGR demonstrates that meaningful temporal adaptation can be achieved without deep retraining or opaque neural recurrence\. By encoding drift as explicit graph structure rather than opaque neural state, it makes temporal adaptation both auditable and domain\-transferable\. In an era of rapidly evolving biomedical knowledge and large\-scale AI discovery, such transparent mechanisms offer a path toward sustainable, continually updating scientific knowledge graphs that balance accuracy, explainability, and efficiency\.

## VIConclusion

This paper introduced DATGR, a Drift\-Aware Temporal Graph Rewiring framework that captures biomedical semantic evolution through lightweight, interpretable edge updates rather than full embedding retraining\. By integrating embedding\-based drift estimation with a logistic rewiring rule, DATGR achieves consistent AUROC gains of approximately 0\.066 over a static baseline while maintaining comparable precision\-recall performance, demonstrating that edge\-level adaptation effectively encodes temporal semantic change at substantially lower computational cost than dynamic GNNs or sequential fine\-tuning approaches\.

More broadly, DATGR illustrates that meaningful temporal adaptation need not rely on opaque neural recurrence or expensive retraining\. Its transparent, graph\-based update mechanism makes it well\-suited for deployment in knowledge graph curation, biomedical RAG pipelines, and semantic retrieval systems where both interpretability and timeliness are critical\. In doing so, it offers a principled bridge between symbolic and neural paradigms for evolving scientific text\.

## Acknowledgment

The authors thank Jacqueline M\. Ring, Vice Chancellor and Chief Institutional Research Officer at Rowan University, for her support and guidance throughout this work\. Portions of this manuscript were refined using OpenAI’s ChatGPT to improve clarity and formatting\[[15](https://arxiv.org/html/2607.08490#bib.bib15)\]\. All conceptual content, data analysis, and interpretation were developed by the authors, in compliance with IEEE’s guidance on AI\-generated text disclosure\.

## Code Availability

The implementation and evaluation scripts for DATGR are publicly available at: https://github\.com/sahanakv01123/DATGR\. The repository includes the full Colab notebook, experiment pipeline, and figure\-generation scripts necessary to reproduce the reported results\.

## Published Version

This paper has been accepted and published at the 2026 IEEE Conference on Artificial Intelligence \(CAI 2026\)\[[17](https://arxiv.org/html/2607.08490#bib.bib17)\]\. The final published version is available at:https://doi\.org/10\.1109/CAI68641\.2026\.11536228

## References

- \[1\]K\. W\. Church and P\. Hanks, “Word association norms, mutual information, and lexicography,” inProc\. 27th Annu\. Meeting Assoc\. Comput\. Linguistics \(ACL\), Vancouver, BC, Canada, pp\. 76–83, 1990\.
- \[2\]T\. Mikolov, K\. Chen, G\. Corrado, and J\. Dean, “Efficient Estimation of Word Representations in Vector Space,” arXiv preprint arXiv:1301\.3781, 2013\. Available: https://arxiv\.org/abs/1301\.3781
- \[3\]J\. Pennington, R\. Socher, and C\. D\. Manning, “GloVe: Global Vectors for Word Representation,” inProc\. Conf\. Empirical Methods in Natural Language Processing \(EMNLP\), Doha, Qatar, pp\. 1532–1543, 2014\. Available: https://aclanthology\.org/D14\-1162
- \[4\]W\. L\. Hamilton, J\. Leskovec, and D\. Jurafsky, “Diachronic Word Embeddings Reveal Statistical Laws of Semantic Change,” inProc\. 54th Annu\. Meeting Assoc\. Comput\. Linguistics \(ACL\), vol\. 1 \(Long Papers\), pp\. 1489–1501, 2016\. Available: https://aclanthology\.org/P16\-1141
- \[5\]N\. Tahmasebi, L\. Borin, and A\. Jatowt, “Survey of Computational Approaches to Lexical Semantic Change Detection,” inComputational Approaches to Semantic Change, N\. Tahmasebi, L\. Borin, A\. Jatowt, Y\. Xu & S\. Hengchen, Eds\., Berlin: Language Science Press, 2021, pp\. 1–91\. Available: https://langsci\-press\.org/catalog/view/303/3027/2374\-1
- \[6\]R\. Bamler and S\. Mandt, “Dynamic Word Embeddings,” inProc\. 34th Int\. Conf\. Mach\. Learn\. \(ICML\), PMLR vol\. 70, pp\. 380–389, 2017\. Available: https://arxiv\.org/abs/1702\.08359
- \[7\]D\. N\. Nicholson, F\. Alquaddoomi, V\. Rubinetti, and others, “Changing word meanings in biomedical literature reveal pandemics and new technologies,”BioData Mining, vol\. 16, article 16, 2023\. Available: https://doi\.org/10\.1186/s13040\-023\-00332\-2
- \[8\]S\. M\. Kazemi, R\. Goel, K\. Jain, I\. Kobyzev, A\. Sethi, P\. Forsyth, and P\. Poupart, “Representation Learning for Dynamic Graphs: A Survey,”Journal of Machine Learning Research, vol\. 21, no\. 70, pp\. 1–73, 2020\. Available: https://www\.jmlr\.org/papers/v21/19\-447\.html
- \[9\]A\. Arnaiz\-Rodriguez, A\. Begga, F\. Escolano, and N\. Oliver, “DiffWire: Inductive Graph Rewiring via the Lovász Bound,” inProc\. 39th Int\. Conf\. Mach\. Learn\. \(ICML\), PMLR vol\. 162, pp\. 10799\-10810, 2022\. Available: https://arxiv\.org/abs/2206\.07369
- \[10\]Q\. Yuan, S\.\-U\. Guan, P\. Ni, T\. Luo, K\. L\. Man, P\. Wong, and V\. Chang, “Continual Graph Learning: A Survey,” arXiv preprint arXiv:2301\.12230, 2023\. Available: https://arxiv\.org/abs/2301\.12230
- \[11\]H\. Zhou, Y\. Li, H\. Wang, and J\. Xiao, “Learning Temporal Difference Embeddings for Biomedical Text,”Bioinformatics, vol\. 38, no\. 23, pp\. 5253–5262, 2022\. Available: https://academic\.oup\.com/bioinformatics/article/38/23/5253/6747952
- \[12\]Y\. Xiao, D\. Steinecke, A\. R\. Pelletier, Y\. Bai, P\. Ping, and W\. Wang, “Know2BIO: A Comprehensive Dual\-View Benchmark for Evolving Biomedical Knowledge Graphs,” arXiv:2310\.03221, 2023\. Available: https://arxiv\.org/abs/2310\.03221
- \[13\]S\. Jeong, J\. Baek, S\. Cho, S\. J\. Hwang, and J\. C\. Park, “Adaptive\-RAG: Learning to Adapt Retrieval\-Augmented Large Language Models through Question Complexity,” arXiv preprint arXiv:2403\.14403, 2024\. Available: https://arxiv\.org/abs/2403\.14403
- \[14\]D\. Pappas, P\. Stavropoulos, I\. Androutsopoulos, and R\. McDonald, “BioMRC: A Dataset for Biomedical Machine Reading Comprehension,” inProc\. 19th SIGBioMed Workshop on Biomedical Language Processing \(BioNLP\), Association for Computational Linguistics, pp\. 140–149, 2020\.
- \[15\]OpenAI, “ChatGPT,” OpenAI, San Francisco, CA, USA, 2025\. \[Online\]\. Available: https://chat\.openai\.com/\. \[Accessed: Oct\. 29, 2025\]\.
- \[16\]M\. Postiglione, D\. Bean, Z\. Kraljevic, R\. J\. B\. Dobson, and V\. Moscato, “Predicting Future Disorders via Temporal Knowledge Graphs and Medical Ontologies,”IEEE Journal of Biomedical and Health Informatics, vol\. 28, no\. 7, pp\. 4238–4248, 2024\. doi: 10\.1109/JBHI\.2024\.3390419
- \[17\]B\. Vijayakumar and S\. K\. Varadaraju, “Drift\-Aware Temporal Graph Rewiring \(DATGR\) for Adaptive Semantic Modeling in Biomedical Text,” in2026 IEEE Conference on Artificial Intelligence \(CAI\), 2026, pp\. 193–198\. doi: 10\.1109/CAI68641\.2026\.11536228\. Available:https://doi\.org/10\.1109/CAI68641\.2026\.11536228