Kernel fisher discriminants for outlier detection.

Area of Science:

• Machine Learning
• Robust Statistics
• Computer Vision

Background:

• Outlier detection is a fundamental problem in statistics.
• One-class Support Vector Machines (SVM) classifiers are a recent approach.
• Existing methods often require pre-specification of outlier fractions.

Purpose of the Study:

• To bridge the gap between kernelized one-class classification and Gaussian density estimation.
• To develop a model-based approach for outlier detection.
• To enable quantification of deviations from a Gaussian model for identifying atypical objects.

Main Methods:

• Kernelized one-class classification integrated with Gaussian density estimation.
• Establishing an exact relation between these two concepts.
• Utilizing a cross-validated likelihood criterion for unsupervised model selection.

Main Results:

• A method to identify atypical objects by quantifying deviations from a Gaussian model.
• Overcoming the limitation of pre-specifying outlier fractions in one-class approaches.
• Demonstrated effectiveness in detecting atypical objects within image databases.

Conclusions:

• The proposed method provides a robust framework for outlier detection.
• It offers a model-based formalization that enhances outlier identification capabilities.
• Experiments confirm its practical applicability in real-world scenarios like image analysis.