Research On Time-aware One-class Collaborative Filtering

Exploiting implicit feedback for top-K recommendation is a popular problem favored by recommender system researchers.Many algorithms have been proposed to solve this problem.Among these algorithms,item-based one-class collaborative filtering and user-based one-class collaborative filtering are widely adopted by many real-world recommender systems.Model-based one-class collaborative filtering models,e.g.,Bayesian personalized ranking(BPR)and factored item similarity model(FISM),are among the most successful models.However,in real-world scenarios,data is often collected with timestamps.Since the models we mentioned above are not designed for data with time information,many efforts have been made by researchers to develop time-aware one-class collaborative filtering algorithms.In this thesis,we focus on achieving higher recommendation performance through exploiting the time information in data.Firstly,for Bayesian personalized ranking(BPR),we propose TBPR.Specifically,we introduce a time-aware decay function into the objective function of BPR,so that each record in the training set has a different weight.Experimental results show that introducing a decay function into BPR can lead to higher recommendation performance.Secondly,we introduce time-aware ensemble learning(TEL),which in fact consists of several TSL base models.Notice that we change its objective function periodically in order to improve the robustness and recommendation accuracy.Finally,we develop a time-aware sequence-based bidirectional item similarity for similarity measurement,and propose a novel sequential item-based OCCF method.In our empirical studies,we find that the performance of our proposed methods are better than that of the baseline methods in most cases.