A Research On Recommendation Approaches For Implicit Feedbacks

As an effective approach addressing information overloading problem,personalized recommendation has been the hotspot in both industry and academia.According to category of feedbacks,it can be classified into recommendation for explicit feedbacks and for implicit feedbacks.While the former is the mainstream,the prevalence of the latter makes it more valuable for research and application.But the constraints of implicit feedbacks have made great challenge to personalized recommendation.The paper firstly give analysis of chracteristics of implicit feedbacks,then a systematic taxnonomy for state-of-the-art recommendation approaches for implicit feedbacks,including one-class collaborative filtering,L2R-based methods and other approaches with auxiliary information.Then the strength&weakness of these approaches,and various evaluation metrics for implicit feedback oriented recommendation are analyzed.On this basis,experiments are taken on KNN collaborative filtering and one-class collaborative filtering embeded with rank similarity from the perspective of ranking for recommendation,which proved to be helpful in improving performance via comparing to user-based collaborative filtering and weighted matrix factorization algorithms,especially for imporving the ranking of recommendation results.Then paper emphasizes the idea of ”ranking” and proposes a noble model MMR based on ranking optimization,which directly maximizes the MAP metric of recommendation result.With the modeling techniques borrowed from BPR recommendation framework,the MAP metric has been transformed into smoothed version,which leads to the gradient-based optimization method.Furthermore,because of the differences of underlay models,MMR can be expaneded into matrix-factorization-based and neighbor-based algorithms accordingly,named MMR-MF and MMR-KNN.After that,the paper presents a faster learning strategy for sampling to address the computational bottleneck,then verifies its correctness and effectiveness through both theoretical derivation and experimental results.Lastly,the MMR algorithms are compared thoroughly with BPR methods and aforementioned baseline algorithms on real-world datasets,which comes to the conclusion that MMR model outperforms others on comprehensive evaluation metrics.Finally it's proved that L2R-based method with MAP optimization possesses advantages over other recommendation approaches,and feasbility with proper sampling&optimization strategies.