Research And Design Of Optimization Recommendation Algorithm Based On Similarity

In the era of Internet, the personalized recommendation system has been widely used.The recommendation algorithms play a decisive role in the recommendation system, and the mainstream of which are collaborative filtering algorithms at present. One kind of them has improved prediction precision to some extent, which takes advantage of User-based CF to pre-fill the sparse rating matrix, and then predicts the unknown ratings by adopting Item-based CF. However, two problems exist: In the pre-filling stage, the similarity of users is affected by the sparsity of rating data in the calculation process,which leads to the inaccuracy of the nearest neighbor set. As a result, the accuracy of dense data which is obtained by user's nearest neighbor set is low. In the predicting process, due to the irrationality of the process of measuring the similarity of items, the nearest neighbor set of item is not accurate enough. Lastly, Eventually the algorithm accuracy descend.In order to further improve the prediction accuracy of the algorithm, this paper makes optimization for the two issues: In the pre-filling stage, according to a optimization model of similarity matrix proposed, this paper learns its parameters to obtain the matrix which characterizes the user's similarity and more accurate user's nearest neighbor set, and this process is less affected by the sparsity of data. Then this paper makes use of the user's nearest neighbor set to pre-fill the original rating matrix for getting more accurate and dense rating matrix. In the predicting process, this paper takes into account the similarity between the common scoring users and the target users to optimize the measure of similarity between items, so we get more accurate item's nearest neighbor set. Finally, we predict unknown ratings of items by using the item's nearest neighbor set on the base of the dense rating matrix. In this paper, the design of algorithm is based on the two optimization strategies above. At last, this paper carries out multiple sets of contrast experiments by using an open source set of data. The algorithm in this paper could be preciser than the pre-optimized algorithm according to analysis of the experimental result,which verified the validity of the algorithm.