Weighted t-Distributed Stochastic Neighbor Embedding for Projection-Based Clustering

This paper presents a projection-based clustering method for visualizing high-dimensional data points in lower-dimensional spaces while preserving the data’s structural properties. The proposed method modifies the t-Distributed Stochastic Neighbor Embedding (t-SNE) algorithm by adding a weight function that adjusts the dissimilarity between high-dimensional data points to obtain more realistic lower-dimensional representations. In our algorithm, the centroids obtained with a prototype-based clustering algorithm attract high-dimensional data points allocated to their respective clusters, while repelling those points assigned to other clusters. The simulations using real-world datasets show that the Weighted t-SNE produces better projections than similar algorithms without the need for any previous dimensionality reduction step.