PROJECT TITLE :

Saliency-Based Multilabel Linear Discriminant Analysis

ABSTRACT:

The objective of the traditional statistical machine-learning method known as linear discriminant analysis (LDA), which seeks to find a linear data transformation that will increase class discrimination in an optimal discriminant subspace, is to find a linear data transformation. The Gaussian class distribution and the use of single-label data annotations are two of the underlying premises of traditional LDA. In this article, we propose a new variation of LDA that can be utilized in multilabel classification tasks for the purpose of accomplishing Dimensionality Reduction on the original data in order to improve the subsequent performance of any multilabel classifier. A probabilistic approach to estimating the saliency of classes is presented here as a means of computing weights for all instances that are based on saliency. When it comes time to calculate the projection matrix, we make use of the weights to redefine the between-class and within-class scatter matrices that are required. We formulate six different variants of the proposed saliency-based multilabel LDA (SMLDA), each of which is based on a different type of prior information regarding the significance of each instance for their class(es), which is extracted from their labels and features. According to the results of our experiments, the proposed SMLDA is superior to a number of alternative strategies for dimensionality reduction in terms of its ability to improve performance in a variety of multilabel classification problems.