Enhancing support vector machine performance through boundary-focused covariance matrix learning

Support Vector Machine (SVM) has proven effective in various classification tasks by focusing on maximizing the margin between classes. However, standard SVM often fails to consider the underlying geometric structure of data near the decision boundary, particularly in scenarios where class overlap or noise occurs. This paper proposes a new method that focuses on boundary-critical samples those near the decision boundary by integrating their covariance matrix into the learning process. The principal eigenvector of this covariance matrix is then utilized to guide the classifier towards the most informative regions of the data, enabling it to capture the local geometry better. Importantly, this modification is confined to the objective function, ensuring the convexity of the optimization problem is preserved. Experimental results across various datasets, including linearly and non-linearly separable ones, demonstrate that the proposed method provides competitive or slightly improved classification performance.

کلیدواژه ها:

support vector machine ، Boundary Samples ، Covariance Matrix ، Eigenvector