Effective public service delivery supported by time-decayed Bayesian personalized ranking

摘要

Information overload is becoming a prominent problem that hinders the effectiveness of e-government. Personalized e-government services with recommendation techniques may provide a solution by helping users find the target service items based on her interaction histories. Most existing e-government recommendation methods are rating based, but the ratings in e-government are unavailable. Besides, the interactions in e-government are sequence-dependent while the sequential information are usually ignored. These two problems negatively affect the efficiency of the service distribution of e-government platforms. In this paper, we studied the sequential implicit feedback recommendation problem in e-government and design a novel learning algorithm called time-decayed Bayesian personalized ranking. Time-decayed BPR captures the sequence effects using time-decayed Ordinal Utility and inherits the seminal pairwise learning framework of BPR. It digests users’ sequential behavior naturally, and combines the benefits of time-aware information (i.e. sequential behavior) and time-invariant information (i.e. general taste). An empirical analysis of real-world e-government datasets shows that our time-decayed-BPR approach improves the performance significantly regarding various evaluation metrics compared with the state-of-the-art baseline methods.