T-Distributed Stochastic Neighbor Embedding Based on Cockroach Swarm Optimization with Student Distribution Parameters

Abstract

T-distributed stochastic neighbor embedding (t-SNE) is a classical dimension reduction method in manifold learning, as it solve the optimization difficulty during training process and overcrowding problem by comparing with traditional stochastic neighbor embedding (SNE). In the core projection step of t-SNE, in order to realize the similarity of probability distribution before and after projection, gradient descent method (GD) or stochastic gradient descent method (SGD) is generally used to solve the minimum parameter of objective function which is defined by Kullback-Leibler divergence (KL divergence). However, GD and SGD may often fall into the trap of local extremum, which means that their global optimization ability is limited. Till now, research on optimizing the solution process for the objective function in t-SNE can be hardly found. By considering that swarm intelligence based algorithm always have a better performance in finding the global extremum, in this paper, a t-SNE based on cockroach swarm optimization with student distribution parameters has been proposed, based on a thorough deduce and related numerical analysis, the effectiveness of the proposed algorithm in finding the global minimum of KL divergence for t-SNE has been verified.