Temporal Diversity in Recommender Systems

As we explored in the previous chapter, CF algorithms are often evaluated according to how accurately they predict user ratings [HKTR04]. However, as recommender systems grow dynamically, a problem arises: current evaluation techniques do not investigate the temporal characteristics of the produced recommendations. Researchers have no means of knowing whether, for example, the system recommends the same items to users over and over again, or whether the most novel content is finding its way into recommendations. The danger here is that, as results may begin to stagnate, users may lose interest in interacting with the recommender system. In this chapter, we investigate one dimension of temporal recommendations: the diversity of recommendation lists over time. We first examine why temporal diversity may be important in recommender system research (Section 5.1) by considering temporal rating patterns and the results of an extensive user survey. Based on these observations, we evaluate three CF algorithms’ temporal diversity from 3 perspectives (Section 5.2): by comparing the intersection of sequential top-N lists, by examining how diversity is affected by the number of ratings that users input, and by weighting-in the trade-off between accuracy and diversity over time. We finally design and evaluate a mechanism to promote temporal diversity (Section 5.3), comparing its performance to a range of baseline techniques. We conclude in Section 5.4 by discussing future research directions.