Abstract: “Nonnegative Tensor Factorization has previously been used in many multi-way data analyses. We use NTF model to do personalized paper recommendation. For recommendation, we analyze four different multiplicative algorithms for NTF based on different decomposition models and different optimization functions. On one hand, one part of algorithms use CP decomposition, the other part use Tucker decomposition. On the other, half of algorithms minimize the least squares error while the others minimize the Kullback- Leibler divergence. Further, we also compare recommendation performance with different rank NTFs. From our experiments, nonnegative Tucker decomposition based on KL divergence has the better result, and to some extent, lower rank NTF can get most of information from dataset.”

Keywords: personalized recommendation, sparsity, tensor factorization, multiplicative update