Keywords: T-SNE, UMAP, Polynomial Regression, PCA, Visualization
Today's complex applications require nonlinear alternatives to PCA. Here we compare three such methods: t-sne, UMAP and a new method we have developed, prVis. T-distributed Stochastic Neighbor Embedding (t-sne) pairs data points across different dimensions and minimizes relative entropy (Van Der Matten and Hinton, 2008). Uniform Manifold Approximation and Projection (UMAP) is similar to t-sne but can also be used to perform general dimension reduction in non-visual contexts (McInnes, Healy, and Melville, 2018). Polynomial Regression Visualization (prVis) visualizes nonlinear relationships in data by using polynomial expansion in concert with PCA. We will compare the performance and quality of visualizations generated by the aforementioned methods on a variety of data sets and demonstrate the insights that can be gained from the use of these tools.