Agentic data science with the Data Agent Kit
cloud.google.com·8h ago
AI Research
arXiv CS.AI·Sep 4·5 min read
Witnesses Explain Anomalies
TL;DR
WAND is a novel unsupervised anomaly detection model that provides inherent explanations for flagged anomalies by identifying witness directions in feature space. It achieves high interpretability without sacrificing accuracy, outperforming existing methods like SHAP and LIME.
✦ Why It Matters
Implement WAND for anomaly detection to gain clear insights into flagged data points without extra computational costs.
Key Takeaways
- Use WAND for efficient anomaly detection with built-in explanations.
- Leverage witness directions for clear feature attributions.
- Achieve high accuracy without the cost of post-hoc explanations.
- Test WAND on your datasets to evaluate its performance.
How It Works
WAND organizes its computation around directions on the unit sphere, scoring each data point based on how far its projection escapes a sub-Gaussian extreme-value baseline. The model identifies witness directions, which are vectors in feature space that indicate the features contributing to the anomaly score.
This allows for per-feature attribution that is computed simultaneously with the anomaly score, making it efficient and interpretable.