Browsing by Author "Zhao, Jie"

Dimension reduction presents expanding importance and prevalence since it lessens the challenge to data visualization and exploratory analysis that numerous science areas rely on. Recently, nonlinear dimension reduction (NLDR) methods have achieved superior performance in coping with complicated data manifolds embedded in high dimensional space. However, conventional statistic software for NLDR visualization purpose (e.g Multidimensional Scaling) often gives undesired desirable layouts. In this thesis work, to improve the performance of NLDR for data visualization, we study the recently proposed and efficient neighbor embedding (NE) framework and develop its software package in statistic software R. The neighbor embedding framework consists of a wide family of NLDR including stochastic neighbor embedding (SNE), symmetric SNE etc. Yet the original SNE optimization algorithm has several drawbacks. For example, it cannot be extended to other NE objective functions and requires quadratic computation cost. To address these drawbacks, we unify many different NE objective functions through several software layers and adopt a tree-based approach for computation acceleration. The core algorithm is implemented in C++ with an lightweight R wrapper. It thus provides an efficient and convenient package for researchers and engineers who work on statistics. We demonstrate the developed software by visualizing the two-dimensional layouts for several typical datasets in machine learning research including MNIST, COIL-20 and Phonemes etc. The results show that NE methods significantly outperform the traditional MDS visualization tool, indicating NE as a promising and useful dimension reduction tool for data visualization in statistics.