Dimensionality reduction decreases the number of variables while keeping the essential information, in order to visualize, speed up learning, or limit noise. It works by extracting new variables (PCA, t-SNE) or by selecting existing variables. It loses information and sometimes makes interpretation harder.