Supervised dimensionality reduction of proportional data using mixture estimation

In this paper, an effective novel approach for dimensionality reduction of labeled proportional data is proposed. By avoiding formulating an eigenvalue problem and constructing a neighborhood graph, the introduced method mitigates some of the major problems from which the well-known algorithms in this category suffer. These disadvantages include problem handling multi-modal or sparse data as well as curse of dimensionality. The devised method transfers the data from high-dimensional space into low-dimensional space using a linear transform which is optimized using an information theoretic measure. To find this projection, a novel approach has been adopted in which projected data are transfered into the low-dimensional space first, and a mixture of distributions is estimated using the projected data for each class separately. In the next step, the distance between the estimated distributions is used as a measure of separation for data classes, and a heuristic search is carried on to find the optimal projection. The effectiveness of the proposed algorithm is demonstrated using different datasets in different scenarios in comparison with other well-known algorithms in the literature. (C) 2020 Elsevier Ltd. All rights reserved.