Unifying rating-oriented and ranking-oriented collaborative filtering for improved recommendation
We propose a novel unified recommendation model, URM, which combines a rating-oriented collaborative filtering (CF) approach, i.e., probabilistic matrix factorization (PMF), and a ranking-oriented CF approach, i.e., list-wise learning-to-rank with matrix factorization (ListRank). The URM benefits from the rating-oriented perspective and the ranking-oriented perspective by sharing common latent features of users and items in PMF and ListRank. We present an efficient learning algorithm to solve the optimization problem for URM. The computational complexity of the algorithm is shown to be scalable, i.e., to be linear with the number of observed ratings in a given user-item rating matrix. The experimental evaluation is conducted on three public datasets with different scales, allowing validation of the scalability of the proposed URM. Our experiments show the proposed URM significantly outperforms other state-of-the-art recommendation approaches across different datasets and different conditions of user profiles. We also demonstrate that the primary contribution to improve recommendation performance is contributed by the ranking-oriented component, while the rating-oriented component is responsible for a significant enhancement.