Dimensionality Reduction Techniques for Data Visualization

Reference: Reducing Snapshots to Points: A Visual Analytics Approach to Dynamic Network Exploration by Elzen et. al. Best Paper of VAST 2015.

Principal Component Analysis (PCA)
Multidimensional Scaling (MDS)
MDS（multidimensional scaling）多维尺度分析 // 通常MDS可以被看做是一个优化问题
t-Distributed Stochastic Neighbour Embedding (t-SNE)
t-SNE降维方法

The computation times for both PCA and t-SNE can be strongly reduced by using improved variants such as Randomized PCA and Barnes-HutSNE, respectively. This makes them usable for large datasets and enables interactive real-time analysis.

PCA is a linear dimensionality reduction technique, i.e., the resulting dimensions are linear combinations of the original dimensions such that the variance of the data is described best. This restriction can be overcome by applying a kerneltrick to achieve non-linearity.