Learning Decision Boundaries for Multidimensional Anomaly Detection.

Area of Science:

• Data Science
• Machine Learning
• Artificial Intelligence

Background:

• Traditional anomaly detection methods struggle with multidimensional data where outliers manifest differently across dimensions.
• Heterogeneity in discriminative spaces across dimensions makes direct comparison of anomaly scores difficult.

Purpose of the Study:

• To address the challenge of comparing anomaly scores from heterogeneous discriminative spaces in multidimensional anomaly detection.
• To introduce a novel model, ALOE (Maximum Margin Multidimensional Anomaly Detection), for effective outlier identification in complex, multi-dimensional datasets.

Main Methods:

• ALOE formulates a convex optimization problem with nonlinear constraints to learn multiple decision boundaries.
• It employs the maximum margin principle and covariance regularization to distinguish outliers from normal samples.
• An alternating optimization method is used to find optimal decision boundaries for each dimension, capturing inter-dimensional correlations.

Main Results:

• Extensive experiments were conducted on 12 real-world datasets.
• ALOE's performance was compared against 34 existing anomaly detection methods.
• The results demonstrated the superior performance of ALOE in multidimensional anomaly detection tasks.

Conclusions:

• ALOE provides a robust and effective solution for multidimensional anomaly detection.
• The model's ability to capture correlations among dimensions enhances outlier identification accuracy.
• ALOE represents a significant advancement in handling complex, heterogeneous anomaly detection scenarios.