Optimizing Class Separability via Projection-Based Discriminative Basis Selection

İsmail Özçil, A. Bugra Koku, Ali Sekmen, Bahadir Bilgin

Session: General

Keywords: Subspace clustering, subspace segmentation, multidimensional classification, dimensionality reduction

TL;DR: The paper proposes a novel method to enhance class separability in high-dimensional classification tasks by selecting the most discriminative basis vectors using singular value decomposition (SVD) and mean average precision (mAP) analysis.

Abstract: In high-dimensional classification tasks, data from different classes often lie in a union of lower-dimensional subspaces. Identifying the basis vectors for each subspace that effectively differentiates between classes can enhance the explainability and accuracy of classification methods. This study proposes a novel approach that uses singular value decomposition to identify class-specific basis vectors that maximize the separability of classes. Instead of selecting the most significant n number of basis vectors using traditional heuristics for basis selection, the mean average precision for each basis vector is calculated, and the top-performing n basis vectors are selected. Furthermore, this study extends the methodology by integrating feature vector outputs from two different pre-trained deep learning models, as input for classification evaluation in two different cases. The proposed methodology is validated through simulations, demonstrating its potential for improving classification in high-dimensional spaces.

Submission Number: 96