Anomaly Explanation

Véronne YepmoIJCAI 2022 · DOI

TL;DR

This paper highlights the critical need for explaining anomaly detection systems, arguing that current methods often fall short. The authors propose a taxonomy of explanation types and outline future work focused on building end-to-end, structure-based anomaly explanation systems, particularly for Isolation Forests.

What Problem Does This Paper Address?

Anomaly detection, the task of identifying unusual data points, is crucial across many fields, from fraud detection to medical diagnosis. However, simply flagging an instance as anomalous isn't enough. Users need to understand why a particular data point was flagged to trust the system, debug it, or make informed decisions. This is especially true in unsupervised anomaly detection, where the "normal" behavior isn't explicitly defined by labeled classes. The challenge lies in providing meaningful explanations that go beyond simple feature importance, especially when anomalies can deviate in complex ways.

Methods

The paper's primary contribution is a taxonomy of anomaly explanation methods, categorized into four levels:

  • Explanation by feature importance: Identifies which features contributed most to the anomaly detection.
  • Explanation by feature values: Highlights the specific values of features that made an instance anomalous.
  • Explanation by data points comparisons: Compares the anomalous instance to "prototypes" of regular or anomalous data points.
  • Explanation by structure analysis: Examines the underlying structure of the data (e.g., clusters) to explain anomalies.

The authors note that while feature importance is the most explored, structure analysis offers the most detailed explanations but is the least developed. Their future work aims to build end-to-end methods for structure-based anomaly explanation, specifically by modifying the Isolation Forest algorithm. This modification would involve pseudo-random splits that preserve data clusters, allowing the forest to identify and describe these clusters, as well as local anomalies within them.

Datasets

Not discussed in the paper. The authors mention using "artificial and real world datasets" for preliminary results but do not specify them.

Key Insights and Ideas

  • Anomaly explanation is crucial for trust and utility: Simply detecting anomalies isn't sufficient; understanding why is essential for users to trust and act upon the system's output.
  • Feature importance is often insufficient: While a common approach, feature importance alone may not capture the full nuance of why an instance is anomalous, especially in complex datasets.
  • Structure analysis offers deeper insights: Explaining anomalies in the context of data structures (like clusters) provides more comprehensive and actionable explanations.
  • The lack of end-to-end methods is a significant gap: Current structure-based explanation methods often rely on pipelines, suggesting a need for integrated solutions.
  • Isolation Forests are promising for structure-based explanation: Their inherent tree structure can be adapted to reveal data substructures and explain anomalies within them.

Key Results

The paper presents preliminary results on artificial and real-world datasets indicating that their modified Isolation Forest approach for structure-based anomaly explanation is "really encouraging." However, specific quantitative results (e.g., accuracy improvements, explanation quality metrics) are not detailed in this abstract.

Limitations and Future Work

The authors acknowledge that anomaly explanation is an under-explored field, particularly the development of end-to-end methods for structure analysis. Their primary future work direction is to build such a method using a modified Isolation Forest. They also highlight the need for better evaluation metrics for anomaly explanations, as current methods often rely on unavailable ground truths. They are developing objective metrics to address this gap.

Summary updated

BibTeX
@inproceedings{yepmo2022anomaly,
  author = {Véronne Yepmo},
  title = {{Anomaly Explanation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.24963/ijcai.2022/844},
}
Figure