User interest dynamics on personalized recommendation

Four real recommender system datasets, the Netflix, SMovieLens, LMovieLens and RYM datasets, are analyzed to gain an insight into their user interest characteristics. A preference of active users to cold objects and a diverse interest of inactive users are revealed, which characteristics are introduced to improve the personalized recommendation algorithms. Based on seven different algorithms, we propose a general improvement formula for them, and finally four new algorithms are obtained. Tested on the four datasets, all the new algorithms are found to outperform the original ones in recommendation accuracy, diversity and novelty, except for the diversity and novelty compared with a heat conduction algorithm. And the recommendation accuracy for the cold objects, referring to the objects with small degrees, is also improved. Moreover, one of the new algorithms shows better performance than two other excellent algorithms in many aspects, i.e., the hybrid algorithm of heat conduction and mass diffusion, and the biased heat conduction algorithm. Our work may shed a new light on personalized recommendation from the perspective of connecting empirical study with algorithm design.