Adaptive Collaborative Filtering for Recommender System

On online websites or e-commerce services, the explosive growth of resource makes the problem of content exploring increasingly challenging. The recommender system is a powerful information filtering tool to support user interaction and promote products. Dealing with determining customer interests, graph-based collaborative filtering is recently the most popular technique. Its only drawback is high computing cost, leads to bad scalability and infeasibility for large size network. Moreover, most previous studies concentrate solely on the accuracy of user preference prediction, while the efficiency of recommendation methods should be considered in many characteristics with complicated relationships, depending on particular systems: popularity, diversity, coverage, congestion. Attempt to conquer these challenges, we propose Adaptive Weighted Conduction formula handling multiple metrics, then construct a scalable model with small complexity, named Adaptive Collaborative Filtering. Experiments are conducted on Movielens, a public dataset, and FPT PLAY, a dataset of our media service. We have an increase of 6% on precision and get close to the best of previous methods on diversity, coverage and congestion. This result shows that the proposed model automatically reveals and adapts to varied requirements of recommender systems, reaches high accuracy and still balances other evaluation aspects.