Regularized Supervised Distance Preserving Projections for Short-Text Classification

Short-text classification is a challenging natural language processing problem. Beyond classification accuracy, another issue refers to the dimensionality of the feature vectors used for classification. This is especially important for embedded applications with hard constraints of computational power and memory. To deal with such problems, many techniques of dimensionality reduction have been developed over the last years. The Supervised Distance Preserving Projections (SDPP) has shown promising results. This work proposes a modified version of the SDPP method, called Regularized SDPP, which relies on the regularization theory. On the basis of experimental evaluation, the proposed approach has achieved good results in comparison to the state-of-the-art methods in nonlinear dimensionality reduction.