Tag
该论文研究了重建式时间序列异常检测中的测试时适应(TTA),发现即使仅使用基准中标注为正常的样本进行更新,也不能可靠地提升检测性能,表明仅改进样本选择并不能完整解决重建式TSAD的TTA问题。
GreptimeDB v1.3.1 is a maintenance release fixing data loss, WAL sync, security, query correctness, and performance issues, plus a Dashboard update adding a Service Map to the Traces view.
IBM and Confluent are making IBM's time-series foundation models — with over 50 million downloads — available natively inside Confluent's Streamhouse via Apache Flink SQL, enabling real-time forecasting, anomaly detection, and optimization on CPUs in the cloud and on premises.
The paper proposes PACT (Piecewise Adaptation via Chart Transport), a method that adapts pretrained time-series foundation models to new operating conditions by partitioning the frozen embedding space into local charts and learning gated residual transports, achieving macro-F1 above 0.95 with only 5% labeled calibration data and no condition identity at test time.
GreptimeDB v1.4 will support ingesting InfluxDB Line Protocol, OTLP, and Prometheus Remote Write payloads via PostgreSQL's COPY FROM STDIN, merged into nightly builds for unified bulk ingestion over a single SQL connection.
The paper proposes a process mining-enabled federated anomaly detection pipeline that uses similarity metrics against a global node to dynamically weight local models during aggregation, addressing non-IID data heterogeneity in industrial IoT. Experiments on manufacturing plant time-series data show 46.52% and 11.99% F1-score improvements over isolated training and FedAvg, plus a 51.24% reduction in false positives.
The paper introduces a synthetic benchmark for target-centric causal discovery in industrial batch processes, built from a two-level structural causal model whose components let difficulty be tuned one condition at a time. Eight methods are evaluated with error analysis showing they degrade in ways consistent with their assumptions as difficulties combine.
This paper proposes an availability-aware historical feedback learning framework for predicting electric arc furnace endpoint temperature and the final concentrations of twelve chemical elements, integrating eight industrial data sources and evaluating twelve regression algorithms. The historical-feedback representation reduces RMSE in 147 of 156 comparisons, with ExtraTrees achieving a 5.54% RMSE improvement for temperature prediction.
The paper introduces WETW, a learned pair-specific weighted Euclidean distance for efficient Dynamic Time Warping (DTW) similarity search, plus the WESee encoder, achieving 11.4x speedup over exact baselines with 0.93 mean recall@10 across eight datasets.
FlexCast 提出一种灵活的上下文-目标学习框架,将时间序列基础模型的可变上下文条件化与自回归分解解耦,通过动态采样可见上下文和掩码目标的掩码预测实现统一的掩码-预测接口,在保持有竞争力精度的同时显著降低推理延迟,并表明自回归模型的关键收益来自灵活训练而非顺序推理。
The paper proposes QUAL, a label-free framework that aligns quantiles of time-series foundation models using the model pool as a probability reference, then applies Huber aggregation for robust zero-shot forecasting, reducing average member MASE by 7.4% and CRPS by 9.8% on GIFT-Eval.
A paper proposing the Task-Aware Diffusion Refiner (TADR), a conditional diffusion model trained against a frozen downstream anomaly detector to fix imputation bias in block-missing IIoT sensor data, improving F1 by 0.338 on SMD while running at 41.4 ms per window on a single CPU core.
论文提出 ATEMPO,一种基于扩散模型的转移熵估计器,可在滞后维度和时间维度上进行摊销估计,单一拟合即可恢复完整的转移熵曲面,并在时变滞后系统、天气预测和脑机接口任务中实现更优的滞后检测与下游预测性能。
The paper introduces the Clustering Occupation Kernel (CLOCK) method, a representation-based framework that embeds time series from dynamical systems into an RKHS to enable clustering via maximum mean discrepancy between attractors, validated on synthetic data and intracranial EEG recordings.
该论文提出了多变量时间序列超分辨率预测(MTS-SRF)任务及 TimeMagnifier 模型,能够从粗粒度历史数据中预测细粒度未来,在十个数据集上以 12.3% 的平均 MSE 降低击败 15 个基线。
该论文提出一个统一评估框架,使用 Jev 1.13.0 作为案例研究 Structured System 1 模型在时间序列分类、事件类型预测、油温方向及股价走势预测等任务上的表现,结果显示其在显式时间模式和事件预测上较强,但在许多数值序列分类和股票走势预测上接近随机水平。
A preregistered causal audit finds that covariate-aware time series foundation models (Chronos-2, TiRex-2, TabPFN-TS, TimesFM-2.5) forecast associations rather than intervention effects, with a derived confounding gap that can be reduced up to 83% via brief interventional post-training with LoRA adapters.
Timestral is a family of four time series foundation models (5M to 209M parameters) achieving state-of-the-art probabilistic forecasting on GIFT-Eval, TIME, and fev-bench via decoupled horizon forecasting, native future covariate support, and synthetic pretraining followed by mixed post-training.
IBM Research 发布 Granite Timeseries Ensemble R1,将四个互补的开源 Granite 时序预测模型组合成一个统一接口,提供概率与分位数预测,无需重训练即可作为默认选择使用。
AnoMix 提出一种多教师排序蒸馏方法(CoRank),将多个异构无监督时间序列异常检测器蒸馏为单一轻量学生模型(<80k 参数),在七个基准数据集上优于单个检测器和静态分数集成,其零样本变体还能在未见数据集上超越大型时间序列基础模型。