Where and How: Mining Convertible Outlying Aspect for Outlier Interpretation.

Area of Science:

• Data Mining
• Machine Learning
• Anomaly Detection

Background:

• Outlier interpretation methods are crucial for understanding data anomalies.
• Existing methods identify outlier subspaces but lack personalization and completeness.
• Personalized interpretation is needed for actionable insights into outlier behavior.

Purpose of the Study:

• To propose a novel convertible outlying aspect mining method, META, for outlier interpretation.
• To identify a personalized outlying feature subspace and quantify the outlier's deviation (how and where).
• To provide actionable interpretative insights for understanding outliers.

Main Methods:

• Developed a convertible outlying aspect with a convertible cost to transform outliers into inliers.
• Utilized a pretrained adversary to validate instance inlier status.
• Formulated an objective function to minimize conversion cost, ensure inlier conversion, and reduce subspace size.
• Learned an optimal convertible outlying aspect using the objective function.

Main Results:

• META identifies personalized outlying feature subspaces, outlying direction, and degree.
• The method generates a converted instance that is classified as an inlier.
• Experiments on real-world and synthetic datasets show META outperforms state-of-the-art baselines.

Conclusions:

• META provides a novel approach to outlier interpretation with personalized insights.
• The method offers actionable understanding of outliers through quantifiable deviation.
• META demonstrates superior performance compared to existing state-of-the-art methods.