Research Of Recommendation Algorithm For One-Class Collaborative Filtering

Collaborative Filtering methods have been regarded as the core technologies for recommendation.Researches on Collaborative Filtering pay more attention to explicit feedback.Such data have specifically numerical ratings.We can determine whether data are positive feedback or negative feedback according to ratings.However,explicit feedback is not always easy to obtain.On the contrary,there are lots of data in one-class feedback form,e.g.,"plays" on music,"views" on news.Such data only have the distinction that whether users have observed or not observed them.Observed items indicate that users like them and are positive-feedback data,while most data are unob-served data.Therefore,the challenge of One-Class Collaborative Filtering problem are usually lacking negative feedback,which makes many existing algorithms not directly applicable to the problem.The current mainstream approaches are pairwise methods which assume that users prefer observed items to unobserved items.Some unobserved items may be users' pref-erence,so pairwise preference assumption that all the observed items should rank higher than all the unobserved items does not always hold.Besides,existing pairwise methods may not perform well in terms of Top-N recommendation as they are not rank-biased methods.In this thesis,we refine the pairwise assumption,which is based on pairwise comparisons via maximizing the likelihood of preference difference to further explore the implicit relationship in observed items and unobserved items.Our algorithm can recommend a better satisfied personalized list for users than previous methods.Some researchers also try to solve One-Class Collaborative Filtering problem with listwise thinking.However,due to their difficulty in modeling the inter-list loss and low efficiency on large-scale dataset,listwise methods are not widely used compared with pairwise methods.This thesis proposes a hybrid Listwise-Pairwise Recommendation algorithm which optimizes a lower bound of a well-known rank-bias metric.The new algorithm can recommend a better sequential list for users without increasing the com-putation complexity.In addition,our approaches are basic methods and have extensive applicability.When combining our models with aforementioned works,better performance can also be achieved.Empirical studies demonstrate that our algorithms outperform the state-of-the-art methods on various real-world datasets.