ReTabAD: A Benchmark for Restoring Semantic Context in Tabular Anomaly Detection

ReTabAD introduces a new benchmark dataset for tabular anomaly detection that incorporates semantic context through textual metadata, addressing a gap where existing datasets lack domain knowledge. The research provides 20 enriched datasets, implementations of classical and LLM-based detection algorithms, and demonstrates that semantic context improves both detection performance and interpretability.

ReTabAD tackles a fundamental limitation in tabular anomaly detection research: the absence of semantic context in benchmark datasets. While anomaly detection in tabular data is critical for fraud detection, system monitoring, and financial analysis, most existing benchmarks strip away the textual metadata and domain knowledge that practitioners use to define what constitutes an anomaly. This research gap has constrained model development and prevented algorithms from leveraging the full information landscape available in real-world deployments.

The benchmark addresses this by curating 20 tabular datasets enriched with structured textual metadata including feature descriptions and domain-specific context. Alongside the datasets, the researchers implement multiple detection approaches ranging from classical statistical methods to contemporary deep learning and LLM-based techniques. The zero-shot LLM framework is particularly noteworthy, as it enables context-aware detection without requiring task-specific training, lowering barriers for practitioners to deploy semantically-informed anomaly detection systems.

For the machine learning and data science community, ReTabAD establishes a new standard for how benchmark datasets should be constructed. By demonstrating that semantic context meaningfully improves detection performance and interpretability, the work validates what domain experts have long understood empirically. This has direct implications for enterprise deployments where explainability and accuracy are equally critical.

Looking ahead, this benchmark will likely accelerate research into multimodal anomaly detection systems that effectively integrate textual and numerical information. Future work may explore how different types of semantic metadata contribute to detection performance, and whether automated metadata generation could extend these benefits to unstructured or legacy datasets lacking documentation.

• →ReTabAD provides 20 semantic-enriched tabular datasets addressing the lack of domain context in existing anomaly detection benchmarks.
• →Textual metadata and feature descriptions demonstrably improve both detection accuracy and model interpretability in tabular anomaly detection.
• →A zero-shot LLM framework enables effective context-aware anomaly detection without task-specific fine-tuning.
• →The benchmark implements state-of-the-art algorithms spanning classical, deep learning, and LLM-based approaches for systematic comparison.
• →Semantic context enables domain-aware reasoning, making detection systems more aligned with real-world operational requirements.