Personalized recommendation via network-based inference with time

Wide attention has been given to the bipartite network-based recommendation method because of its higher accuracy, better diversity and lower computational cost than the global rank method and the standard collaborative filtering method. However, prior studies on bipartite network treated equally the ratings obtained at different time windows, which clear mismatches with the practical situation because the items collected recently should maintain more significance than those collected long ago. Besides, time impact has not yet been systematically studied as an essential context with the consideration of user preference drift in bipartite network. This paper proposes a personalized recommendation method named network-based inference with time (NBIt). We process time information firstly by mapping the ratings in short-time windows to long-time windows. Then, a suitable time attenuation function is selected to ensure a real reflection of user preference. And then, we set the initial resource and attractive power of network. Finally, the recommendation process is elaborated. To avoid the risk of optimized bias and over-fitting, we employ the triple division technique to optimize the long time window parameter and the attraction power parameter. Experimental results from two benchmark datasets of different scales show that the proposed NBIt algorithm surpasses the other five representative and advanced bipartite network recommendation methods in both accuracy and personalization. Furthermore, the proposed NBIt method can be used as a framework in which other network-based recommendation algorithms along with their variants are run with accuracy and diversity likely to be improved.