Research On Similarity Measures In Neighborhood-based Collaborative Filtering Algorithm

With the rapid development of Web2.0, the problem of information overload following information underload becomes the top issue when users get access to online information.Recommendation system is created as one important tool to solve the issue, and the effective recommendation algorithms are the core parts of the whole system.In recent years, collaborative filtering recommendation algorithm has attracted people’s attention in many fields for its unique advantages. Nearest neighborhood based collaborative filtering recommendation algorithm is the most intuitive one, whose construction is primarily built upon two important factors:the similarity measure and the neighborhood selection. There is an implicit assumption in neighborhood-based filtering method,which says that two users who are similar to each other, their ratings to the same item tend to be similar;the stronger their similarities are, the tendency is greater.Two items which have stronger similarity relationship are more likely to get a similar rate from certain user.Similarity measures the correlation between users and between items, and can be calculated based on the rating matrix of users to items and other important characteristics of users and items. Besides,it is also the basis to construct neighborhood of users and items.To accurately measure the similarity between users and items and create neighborhood, then make recommendation for users,it is a vital component to design an appropriate similarity in a neighborhood-based filtering recommendation algorithm.In this work, eight popular similarity measures were chose as the baseline models to compare with each other in terms of recommendation accuracy. We proposed a hybrid model based on some of these baselines to build strongly coherent neighborhoods, the experimental results on Movielens indicates that the similarity measure and the neighborhood selection greatly affect the recommendation performance.The proposed hybrid model could obviously improve the stability and the accuracy of the recommendation algorithm.