Research On Information Similarity In Recommender Systems And Its Applications

Recommender systems have proven to be an important solution to theinformation overload problem, by providing users with proactive andpersonalized information. And collaborative filtering techniques have provento be a vital component of many such recommender systems. The core issue isfinding really similar neighbors. Thus, similarity is the core issue.In this paper, some key technologies including design and architecture ofrecommender systems, design of core recommender engines, similaritymeasurement and so on will be investigated in detail. The contribution of thispaper is mainly reflected in the following three aspects:Firstly, design a common architecture for recommender system, whichcombines various recommender engines for different type of data.Secondly, the popular similarity measures in the current CF will be analyzedtheoretically, and the limitations will be investigated theoretically for somespecific similarity measures. Generally popular similarity measures includecosine-based similarity, adjusted cosine-based similarity, Pearson correlationcoefficient-based similarity, and etc. These measures focus on the linear cases. When the number of neighbors is small, the predicted rating obtained by thesemeasures is not acceptable.Finally, this paper provides three new mutual information(MI)basedsimilarity and an in-depth analysis of shilling attract and presents a novel robustcollaborative filtering system which protect from shilling attract. In the end, theresult of the experiments show this system can lead to improved detectingaccuracy experimentally. Experiments show that the measure can improve therecommendation accuracy.