Interpretable Anomaly Detection and Root Cause Localization

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2442 papers541 conceptsfrozen 2026-10-04

In high-stakes environments like semiconductor manufacturing, cloud infrastructure, and medical diagnostics, knowing that something is wrong is only half the battle. A binary flag or a numerical anomaly score—the standard output of most deep learning models—is often unactionable. If a cloud storage system slows down, an engineer needs to know which of the thousands of metrics is the "culprit." If an industrial sensor flags a defect, a technician needs to know if it's a mechanical failure or simple sensor drift.

The research direction emerging from these papers moves beyond detection toward interpretation and localization. This shift transforms anomaly detection from a "black box" alert system into a diagnostic tool that identifies root causes and provides human-readable justifications.

Moving from Correlation to Causality

Traditional anomaly detection identifies statistical outliers—points that deviate from a learned distribution. However, statistical deviation does not always imply a system failure, and it rarely explains why the failure occurred.

To bridge this gap, researchers like Han Xiao and Shuhan Yuan have introduced frameworks like RootCLAM in On Root Cause Localization and Anomaly Mitigation through Causal Inference. Instead of just looking at raw data, these methods use Structural Causal Models (SCMs) to understand the underlying data-generating process. By modeling the causal relationships between variables, the system can perform "counterfactual reasoning": it asks what the data should have looked like if a specific component hadn't failed. This allows the model to recommend the minimum actions needed to mitigate the anomaly.

Similarly, in cloud computing, Lili Georgieva and colleagues developed the Cloud Causality Analyzer (CCA) in CCA: An ML Pipeline for Cloud Anomaly Troubleshooting. By combining dimensionality reduction with causal inference, CCA can pinpoint performance "culprits" in massive telemetry datasets, providing engineers with a clear path to resolution.

Interpretable by Design: Concepts and Features

A second major strategy involves building models that are "interpretable by design" rather than trying to explain a black box after the fact.

In Transparent Anomaly Detection via Concept-based Explanations, Ivaxi Sheth and Laya Rafiee Sevyeri propose ACE. Instead of learning abstract latent vectors, ACE forces the model to learn human-understandable "concepts" (like "cell density" in medical images). When an anomaly is detected, the model can explicitly state which concept was violated.

For tabular data and tree-based models, researchers have focused on Feature Importance. Mattia Carletti and Gian Antonio Susto introduced DIFFI in Interpretable Anomaly Detection with DIFFI: Depth-based Feature Importance for the Isolation Forest. By analyzing the structure of an Isolation Forest, DIFFI identifies which specific features (e.g., "voltage" or "temperature") were most responsible for isolating a data point as an outlier, providing local explanations that are as fast as the detection itself.

The Rise of Language and Multimodal Explanations

The most recent evolution in this field leverages the reasoning capabilities of Large Language Models (LLMs) to provide natural language feedback. This is particularly useful in industrial settings where operators need qualitative descriptions of faults.

In Large Language Models can Deliver Accurate and Interpretable Time Series Anomaly Detection, Chaoyun Zhang and Chetan Bansal demonstrate LLMAD. This framework doesn't just output a score; it classifies the anomaly type and provides a human-readable explanation of the severity and cause.

Taking this a step further, Jiaxin Zhuang introduced TAMA in See it, Think it, Sorted: Large Multimodal Models are Few-shot Time Series Anomaly Analyzers. TAMA treats time series data as a visual task, allowing a multimodal model to "see" the patterns and "think" through the logical steps of why a specific segment is anomalous. This "See it, Think it, Sorted" pipeline mimics human expert reasoning, making the system's output immediately intuitive to operators.

Visual Analytics and Human-in-the-Loop

Finally, because anomalies are often subjective or context-dependent, a significant body of work focuses on Visual Analytics. These tools don't try to replace the human expert but rather empower them.

Zhuochen Jin and Nan Cao developed EventThread3 in Visual Anomaly Detection in Event Sequence Data, which allows users to compare anomalous sequences against a "normal" baseline visually. Similarly, AnoViz, presented by Patara Trirat and Jae-Gil Lee in AnoViz: A Visual Inspection Tool of Anomalies in Multivariate Time Series, provides variable-wise contribution scores. These visual interfaces turn the "black box" into a "glass box," allowing experts to verify the model's logic and perform their own root-cause analysis.

By combining causal inference, concept-based learning, and natural language reasoning, this research direction ensures that anomaly detection is not just a warning light, but a comprehensive diagnostic engine.

Papers in this version

Showing 60 of 442 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 435 papers
  2. Version 1 · 2026-10-04 · 442 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_black_box_anomaly_detection_explainable__69705f31,
  title = {black-box anomaly detection (explainable anomaly detection)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/black-box-anomaly-detection-explainable-anomaly-detection/rdr_69705f3155a64c45923194fb49cd9234},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_interpretable_anomaly_detection_and_root_69705f31_v1,
  title = {Interpretable Anomaly Detection and Root Cause Localization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/interpretable-anomaly-detection-and-root-cause-localization/rdr_69705f3155a64c45923194fb49cd9234/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}