Hybrid Dimension Reduction Method Based on Isomap and t-SNE with Beetle Antennae Search Algorithm

Abstract

The emergence of dimension reduction algorithm can effectively reduce calculation time, storage space for input and parameters, and can solve the problem of sparse samples in high-dimensional space, thus it has been applied widely. As two typical nonlinear dimension reduction algorithms, isometric feature mapping (Isomap) and t-distributed stochastic neighbor embedding (t-SNE) are also called manifold learning, even if they can realize dimension reduction, both of them have a common disadvantage that they can only find the local optimal solution. Thus, it is of great importance to overcome this shortcoming. In this paper, the two manifold learning methods Isomap and t-SNE have been mixed to form a novel method, which has a totally new loss function in dimension reduction; moreover, beetle antennae search (BAS) algorithm has also been introduced into the proposed method, which has good global convergence, great randomness, and can solve the problem of effectively finding the global optimal solution out.