Recommendation Algorithm Based On Item Similarity For Cold-users

Traditional collaborative filtering recommender system based-on item choosesneighborhood for the target user relies on item similarity. But several existed andcommonly used traditional similarity methods have certain disadvantages. Because thenumber of cold-users rated is small, it is difficult to find the items cold-users rated together.The methods existed can’t compute similarity accurately, solving the cold-start problempoorly. In reaction to the phenomenon, we propose a two stage item similarity algorithmand an improved prediction score algorithm, they improve the accuracy of the calculatedsimilarity and the efficiency of recommendation for cold-users. We mainly focus on thefollowing aspects.Firstly, aiming at cold-start problem, because the classic pearson similarity algorithmcan not calculate similarity between items accurately, we propose a two stage itemsimilarity algorithm. It consists of three parts:Similarity1,Similarity2,Similarity3，and itchooses formula dynamically to compute similarity between items depending on thecomparison of user-item rates and the middle rate. Middle rate is the average of themax-rate and the min-rate.Secondly, In order to improve the accuracy of the prediction score leading to moreaccurate recommendation result, we propose an improved predicting algorithm. First, Ituses the existed predicting algorithm to predict the rates of items that cold-users have rated,and then compute the average of deviation between the pre-predicting rate and the ratecold-users have given to the item, at last it combines the average of deviation and theexisted predicting algorithm to predict the rate of the items that cold-users have not ratebefore.Finally, we test the feasibility and the accuracy for the item-based similarityalgorithm and the improved predicting algorithm through experiments, then based on theexperimental results, we have improved the algorithm.