Fast Gradient Computation for Learning with Tensor Product Kernels and Sparse Training Labels

Supervised learning with pair-input data has recently become one of the most intensively studied topics in pattern recognition literature, and its applications are numerous, including, for example, collaborative filtering, information retrieval, and drug-target interaction prediction. Regularized least-squares (RLS) is a kernel-based learning algorithm that, together with tensor product kernels, is a successful tool for solving pair-input learning problems, especially the ones in which the aim is to generalize to new types of inputs not encountered in during the training phase. The training of tensor kernel RLS models for pair-input problems has been traditionally accelerated with the so-called vec-trick. We show that it can be further accelerated by taking advantage of the sparsity of the training labels. This speed improvement is demonstrated in a running time experiment and the applicability of the algorithm in a practical problem of predicting drug-target interactions.