Study On Data-driven Collaborative Filtering For Implicit Feedback

With the development of information technology,the explosion of information gives people a more convinent life.However,the problem of “information abundance” aslo arises.How to help people find the most interesting information is a key problem.Search enginees can help people to filter out the most useless information.But search enginees just accept and excute the quary.They can't work by themselves and they can't provide the personalized services.As a result,as another method for information filtering,recommender systems have been developed very well.The input of recommender systems has two types,the first is explicit ratings and the other is implicit feedback.Compared with explicit ratings,implicit feedback doesn't need to involve users.As a result,implicit feedback is more common in people's lives.This paper focuses on the study of data-driven collaborative filtering in implicit feedback.The main contribution is shown as follows.1.This paper proposes a recommender algorithms based on data-driven hybrid similarities.The proposed models take both similar users and similar items into consideration.What's more,it utilizes a learning process to learn the similarity directly from data.2.This paper investigates the influence of different loss functions,including square loss function,pairwise loss function and logistics loss function.This paper also takes different regularization constraints into consideration.3.This paper carries out experiments in public datasets to evaluate the proposed methods.Based on the research,a prototype of moive recommendation is implemented.This paper first introduces the research status of recommender systems.Then the paper clarifies the meaning and the background of the research.After that,this paper reviews the related research about recommender systems and figures out the advantages and disadvantages of existing approaches.Then,the approach is described and the experimental result is presented.Finally,this paper investigates the future work about recommender systems and figures out some possible research points.