A small sphere and large margin approach for novelty detection using training data with outliers.

Area of Science:

• Machine Learning
• Data Science
• Computer Science

Background:

• Novelty detection is crucial for identifying anomalies in datasets.
• Existing methods often struggle with imbalanced datasets containing mostly normal and few abnormal examples.
• Outlier detection requires robust algorithms to distinguish between normal and anomalous data points.

Purpose of the Study:

• To propose a "small sphere and large margin" approach for novelty detection.
• To develop a method that effectively handles datasets with a majority of normal instances and a minority of outliers.
• To create a tight and closed boundary around normal data points.

Main Methods:

• Constructing a hypersphere encompassing the majority of normal training data.
• Minimizing the hypersphere's volume while maximizing the margin to outlier data.
• Utilizing convex optimization, solvable with nu-Support Vector Machines (ν-SVM) software.

Main Results:

• The proposed method establishes a closed and tight boundary around normal data.
• Experimental results demonstrate the algorithm's effectiveness in novelty detection.
• The approach successfully differentiates normal data from outliers even with imbalanced datasets.

Conclusions:

• The "small sphere and large margin" approach is an effective technique for novelty detection.
• This method provides a robust solution for anomaly detection in datasets dominated by normal instances.
• The algorithm's reliance on convex optimization ensures efficient computation and practical applicability.