Maximizing generalized mean of trace ratios for discriminative feature learning
Jiyong Oh · Pattern Recognition Letters · 2025
Although linear discriminant analysis (LDA) is effective and efficient, it does not always succeed in obtaining discriminative features. This paper deals with the class separation (CS) problem leading to the failure of LDA. This study first presents a general framework leveraging the generalized mean in which some previous methods to address the CS problem become specific cases. The proposed framework is meaningful because it integrates the previous methods in two different approaches from a single framework. Then, this study proposes a new dimensionality reduction method by employing the trace ratio of between-class scatter and within-class scatter as a dissimilarity measure, which is also a special case of the proposed framework. Classification experiments on five datasets demonstrate that the proposed method is competitive with other state-of-the-art methods. • This paper addresses the class separation problem of linear discriminant analysis. • This paper presents a general framework for approaches to alleviate the problem. • The framework bridges the gap between two different approaches to handle the problem. • This paper proposes a new dimensionality reduction method to overcome the problem. • Experiments show that the proposed method is consistently the most competitive.