Large-scale data visualization with BH t-SNE, LargeViz, and Umap

This work searches the characteristics of the algorithms in real data sets of different instances and dimensions, and uses neighborhood-based DR performance criteria to provide method performance for evaluating the performance of the results. The work also generates heatmap for parameter pairs of each algorithm. By doing this, it discovers how each parameter changes the result and the similarities, as well as the differences between similar parameters in different algorithms and the reasons for the differences in performance between algorithms in the same data set. In the end, the optimal parameters of the three algorithms applied on each dataset with the help of the grid search algorithm. The examining code generates the visualization result with related DR performance score in 2D on each dataset and related time consumption. The applicability of each algorithm in different situations is also summarized.