Divergence Based Non-Negative Matrix Factorization for top-N Recommendations

AbstractPersonalized top-N recommendation algorithms are among the most effective techniques providing customized suggestions in information retrieval applications. Most of the current methods construct personalized recommendations based on various loss functions such as pairwise ranking loss and point-wise recovery loss. In this paper, we propose a personalized top-N recommendation method based on non-negative matrix factorization with divergence as a point-wise ranking loss function. Our method finds the latent factors from the existing data to improve recommendation predictions. We formulate the learning problem with regularized divergence as a constrained non-convex minimization problem and develop a projected gradient descent optimization algorithm to solve the divergence problem. We evaluate our approach using six personal recommendation task related datasets by employing root mean squared error (RMSE) and hit rate (HR). Our experimental results demonstrate improved RMSE and HR for most of the datasets.


Return to previous page