Research On Collaborative Filtering Recommendation Algorithm

With the rapid development of E-commerce and Internet, lots of information make people get lost and the redundancy of information brings many obfuscation and inconvenience for people especially in understanding information and business work, It is necessary to use a personalized recommendation technology to help people resolving kinds of obfuscation. This article researched the relevant technology and concept to the personalized recommendation system, collaborative filtering is the entrance of this article, based on the existing problems in collaborative filtering. We proposed a new resolving way.User-based collaborative filtering algorithm, which is the kernel of collaborative filtering's idea. Which was proposed earliest, and understood easily, has got dominant success especially in the effect of recommendation. However, when the system is very huge, sparsity and cold-start grow serious, which will critically constraint the effect of recommendation, the recommendation system can't obtain the perceiving result.Item-based collaborative filtering algorithm is thinking in a manner of exchanging positions based on traditional collaborative filtering. This algorithm resolved sparsity in part. Even though the matrix is not special sparse, the precision of recommendation is not low compared with user-based collaborative filtering algorithm. But this algorithm can't really resolve sparsity.SVD-based collaborative filtering algorithm uses a method of reducing dimension for the matrix of ratings. Then directly predict the ratings for unrated item. This method deal with sparsity excellently, and the effect of recommendation is excellent, but this algorithm is not easily understood.The character of item constrains the interest of user, so the degree of user loved attribute is relevant to the similarity of item and user. If user has strong interest with the attribute which is belong to item, this is noting influence to the similarity of user or item, on the contrary, this is great influence to the similarity of user or item.Based on the analysis of above, the detail works is shown below:1) The third chapter proposed a Collaborative filtering recommendation algorithm based on decision-making of neighbor, first decomposition the user-item rating matrix, then based on the treated matrix implement close neighbor concept, compute the similarity of users and items, Every item which is waiting for perceiving, compared with the number of user's neighbor and item's neighbor, which is greater then use corresponding algorithm to predict rating. This method introduced the idea of dynamic decision, which is better than along method. The experiment shows excellent effort in recommendation.2) The fourth chapter proposed a Collaborative filtering recommendation algorithm based on bayesian theory predicting item. Every rating item has different attributes first use bayesian theory compute the probability of being liked, incorporate the probability into computing similarity, which means a new method of computing similarity, then use predicting method based on the item, so as to first predict the rating which blongs to the union and not rated by user, then compute the similarity of user based on the union, last compute the resting item's rating which is not rated. Though the value of similarity which has incorporated the influence of attribute favorability. So it reduces sparsity and better conformed to the psychology of people. This method got excellent effort in recommendation, just as showed in the experiments.3) The predicted rating has decimal digits, original rating matrix adopt rank rating, so there is no decimal digits, many research do round for predicted rating, which method can't give expression to the trend of rating. This article implements trade off for predicted rating based on user's trend of rating.4) Through experiment verify above methods, meanwhile improve the effect of recommendation differ in degree. Every method recovered sparsity well. This paper refered to resolving method of cold-start. The problem of sparsity is mainly considered in this paper.