One-Class SVM Hyperparameter Optimization for Class Imbalance

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

In real-life applications, especially in medical analytics like anomaly detection from physiological signals, the scarcity of negative examples due to class imbalance issues hampers effective classification, leading to high misclassification rates and the need for optimal decision boundary construction in one-class support vector machines (OC-SVMs).

Innovation Solution

The method jointly optimizes the kernel coefficient γ and rejection rate hyperparameter v for OC-SVMs to form a smooth, non-linear decision boundary, using a Radial Basis Function kernel and a divide-conquer algorithm to find optimal hyperparameters that minimize misclassification errors and ensure robust anomaly detection.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If traditional OC-SVM is used with single labeled training class, then the classifier can handle class imbalance, but misclassification rate increases due to scarcity of negative examples

Engineering Contradiction:
Improveability to handle class imbalanceVSAvoidmisclassification rate
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent combines multiple hyperparameter optimization techniques (grid search, random search, and Bayesian optimization) into a unified framework that jointly optimizes multiple hyperparameters (C, gamma, nu) simultaneously, rather than optimizing them separately or using single-labeled training approaches

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent systematically varies and optimizes multiple hyperparameters (C, gamma, nu) of the OC-SVM model to find the optimal configuration that minimizes misclassification rate while handling class imbalance, using performance metrics to guide parameter selection

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If domain experts manually label and annotate training datasets, then classification accuracy improves, but the process becomes massively costly and not feasible in many scenarios

Engineering Contradiction:
Improveclassification accuracyVSAvoidlabeling time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent enables the OC-SVM model to automatically optimize its own hyperparameters using performance metrics from validation data, eliminating the need for manual expert intervention in hyperparameter tuning and making the process self-directed and automated

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent implements a feedback mechanism where the model's performance metrics (misclassification rate, anomaly detection accuracy) are continuously evaluated and used to guide hyperparameter optimization, creating a closed-loop system that automatically improves without expert intervention

Inventive Principle:
Principle #23Feedback

3Quantity of substance

If clinically normal training signals are used for medical anomaly detection, then training data availability increases, but the classifier becomes biased and misclassification of abnormal cases increases

Engineering Contradiction:
Improvetraining data availabilityVSAvoidanomaly detection accuracy
Core Design Contradiction:
Quantity of substanceVSReliability

Solution Approach 1:

The patent optimizes the nu hyperparameter which controls the rejection rate and acts as a regularization parameter, preventing the model from being overly influenced by the dominant normal class while still utilizing available training data effectively

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent performs preliminary hyperparameter optimization on validation data before final deployment, adjusting the model's sensitivity to abnormal patterns in advance, which helps prevent bias toward normal cases when the model is applied to actual anomaly detection tasks

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3480734B1Generalized one-class support vector machines with jointly optimized hyperparameters thereof
Publication Date: 2024.02.07 TATA CONSULTANCY SERVICES LTD
  • EP3480734B1 patent drawingFigure 1
  • EP3480734B1 patent drawingFigure 2
  • EP3480734B1 patent drawingFigure 3

AI summary

Absence of well-represented training datasets cause a class imbalance problem in one-class support vector machines (OC-SVMs). The present disclosure addresses this challenge by computing optimal hyperparameters of the OC-SVM based on imbalanced training sets wherein one of the class examples outnumbers the other class examples. The hyperparameters kernel co-efficient y and rejection rate hyperparameter v of the OC-SVM are optimized to trade-off the maximization of classification performance while maintaining stability thereby ensuring that the optimized hyperparameters are not transient and provide a smooth non-linear decision boundary to reduce misclassification as known in the art. This finds application particularly in clinical decision making such as detecting cardiac abnormality condition under practical conditions of contaminated inputs and scarcity of well-represented training datasets.