Collaborative Filtering Algorithm Based On Best Similarity Weight And Localized Preference

Recommendation systems were the core of intelligent e-business, while collaborative filtering was the most successful and most widely used algorithm in recommendation system. But there are some problems including low recommended precision, data sparsity. For these problems, four-part studies made are summarized as follows.1. Aiming at the traditional collaborative filtering(CF) with low accuracy and the existing method with weighting which need prior knowledge or optimal parameter settings and can only improve accuracy from a certain aspect, we propose a collaborative filter recommendation model based on the optimal weighted similarity. This model improves the neighbors’ selection by normalizing similarity of neighbors and uses optimization methods to solve best weight of similarity. This method unifies different weighted algorithm and realizes the best similarity weighted and solution of weight without prior knowledge in theory, besides, it searches for the better parameters by the PSO optimization. Experiment results in MovieLens-100 K data set shows that MAE of collaborative filter recommendation model based on the optimal weighted similarity is lower than the traditional CF, correlation-weight CF, IFUBCF and IFIBCF.2. Aiming at the collaborative filtering with low accuracy and low global similarity in the condition of the sparse data, while some user have common interest in local field, so we proposes a collaborative filter based on the partial neighbors. This algorithm improves similarity calculation in local fields which were divided with the method of item clustering in the way of K-medoids clustering with the initial clusters’ center items chosen the way of minimize the biggest item similarity and the prediction calculation in the cluster. Experiment results in MovieLens-100 k data set and EachMovie data set show that MAE of collaborative filtering based local neighbors is lower than the traditional CF, correlation-weight CF.3. Aiming at the data sparsity, we proposes a collaborative filter based on the fusion of global neighbors and local neighbors. The basic idea is the ratio of the sum of global neighbors’ similar and the sum of local neighbors’ similar to merge the prediction of global algorithm and the prediction of local algorithm, which combine the advantages of the two algorithms. Experiment results in MovieLens-100 k data set and EachMovie data set show that the precision of collaborative filtering based on the fusion of global neighbors and local neighbors is better than the collaborative filtering and collaborative filtering based local neighbors.