Technology Category
Machine Learning: Kernel MethodsData Mining & Knowledge Management: Data CompressionComputer Vision: Learning & Optimization for CV
A Novel Cholesky Kernel based Support Vector Classifier
🤖 AI Summary
Traditional SVMs operating in non-Euclidean spaces neglect second-order statistical properties—namely, variance and covariance—of the data, leading to suboptimal decision boundaries and limited generalization capability. To address this, we propose a novel adaptive kernel function based on Cholesky decomposition of the covariance matrix, which maps input data into a reconstructed Euclidean space where covariance structure is explicitly modeled to optimize margin maximization and decision boundary placement. This work constitutes the first integration of Cholesky decomposition into SVM kernel design, thereby relaxing the theoretical reliance of SVMs on Euclidean metrics and extending their applicability to data distributions with arbitrary covariance structures. Experiments on the Wisconsin Breast Cancer dataset demonstrate that the proposed method significantly outperforms both linear-kernel and RBF-kernel SVMs, achieving substantial improvements in classification accuracy, recall, and F1-score.
Technology Category
Application Category
📝 Abstract
Support Vector Machine (SVM) is a popular supervised classification model that works by first finding the margin boundaries for the training data classes and then calculating the decision boundary, which is then used to classify the test data. This study demonstrates limitations of traditional support vector classification which uses cartesian coordinate geometry to find the margin and decision boundaries in an input space using only a few support vectors, without considering data variance and correlation. Subsequently, the study proposes a new Cholesky Kernel that adjusts for the effects of variance-covariance structure of the data in the decision boundary equation and margin calculations. The study demonstrates that SVM model is valid only in the Euclidean space, and the Cholesky kernel obtained by decomposing covariance matrix acts as a transformation matrix, which when applied on the original data transforms the data from the input space to the Euclidean space. The effectiveness of the Cholesky kernel based SVM classifier is demonstrated by classifying the Wisconsin Breast Cancer (Diagnostic) Dataset and comparing with traditional SVM approaches. The Cholesky kernel based SVM model shows marked improvement in the precision, recall and F1 scores compared to linear and other kernel SVMs.
Problem
Research questions and friction points this paper is trying to address.
Traditional SVM ignores data variance and correlation
Proposes Cholesky Kernel for variance-covariance adjustment
Transforms data to Euclidean space for valid SVM
Innovation
Methods, ideas, or system contributions that make the work stand out.
Cholesky Kernel adjusts data variance-covariance structure
Transforms data to Euclidean space via covariance decomposition
Improves SVM precision, recall, and F1 scores significantly
🔎 Similar Papers
No similar papers found.
