Stochastic Neighbor Embedding with Trust Region Method Combining with Filter

Abstract

Stochastic neighbor embedding algorithm is an important manifold learning algorithm for nonlinear dimensionality reduction in the field of big data and machine learning. In the stochastic neighborhood embedding algorithm, KL divergence is used to measure the approximate degree of the probability distribution of the neighborhood points in high-dimensional space and that of the neighborhood points in low-dimensional space after projection. Gradient descent method or stochastic gradient descent method are often used to solve the minimum value problem of KL divergence. However, gradient descent method and stochastic gradient descent method are easy to fall into the trap of local minimum value. The main purpose of this paper is to give a new kind of stochastic neighbor algorithm, mainly using the trust region method combining with filter to replace the gradient dependent optimization algorithm, so as to get better global convergence, especially for the strong nonlinear problems with faster convergence speed.