Collaborative Filtering Algorithm Based On Neighborhood Relationship

With the continuous development of Internet technology,the Internet has become an indispensable part in people's daily life.However,the issue of information overload on the Internet makes it impossible for people to extract the information they need from mass information.To solve it,recommender system came into being.The main idea behind recommender system is to guess the user's interests and hobbies based on the user's historical behavior data,and recommend matching items based on the guessed hobbies.Collaborative filtering is one of the most widely used technologies in recommender system,and has two critical steps:selecting neighbors for target users and predicting rating for items.However,the behavior data of most users is sparse,which brings challenges to the selection of neighbors and the prediction of items.To deal with the sparse behavior data,this thesis studies the collaborative filtering methods by considering the fast search method and the propagation way of neighbors,which solves the issue of data sparsity well and improves the prediction accuracy.The innovation work of this thesis is summarized as follows:(1)A fast neighbor user searching(FNUS)method for collaborative filtering is proposed.Generally,the combined similarity has a higher recommendation accuracy rate than the simple similarity when applying to the collaborative filtering algorithms.However,the time cost of using the combined similarity to search neighbors is much greater than that of the simple similarity,which reduces the recommendation efficiency.To remedy it,this thesis proposes FNUS for collaborative filtering algorithms.First,according to the user's rating habits,behavior data is divided into three parts:interesting,uninteresting,and neutral,which reflect the spaces of items with different degrees of interests.Then,FNUS selects the nearest neighbors in the each item space,and obtain the indirect neighbors through the neighbor propagation.Finally,the neighbor users in the three item spaces are merged as the final neighbor set of the target user.Simulation experiments on three movie data verify that FNUS can not only improve time efficiency,but also ensure the accuracy of the recommendation to a certain extent.(2)A subspace ensemble-based neighbor user searching(SENUS)method for collaborative filtering is proposed.Generally speaking,fast algorithms would bring some performance loss.In order to improve the performance of FNUS,this thesis proposes SENUS based on subspace integration.On the basis of the three item spaces,this thesis defines a common rating support to calculate the trustworthiness between users.SENUS needs to fuse the three item spaces into a space by weighting,and searches the neighbor user set in this new space.The generated neighbor user set reflects the differences between different item spaces and reduces the influence of invalid neighbor users.Experimental results on three actual data sets show that SENUS has a better recommendation performance for not only providing good neighbors but also costing less time.(3)A Neighborhood-based Iterative Rating Prediction(NIRP)for collaborative filtering is proposed.The issue of data sparsity not only affects the selection of neighbors,but also has a great impact on the prediction of items,which are the two key steps in collaborative filtering algorithms.To further improve the performance of recommendation algorithms,this thesis introduces the idea of neighborhood propagation and proposes a novel prediction algorithm NIRP.The proposed prediction algorithm uses the propagation of neighbors to transfer information,so as to iteratively update the predicted score of items.In the iterative process,NIPR takes into account the reliability of the rating information provided by neighbor users by introducing propagation weights that decay with the number of iterations.Experimental results show that NIRP has better prediction ability,improves the accuracy of recommendation,and better solves the issue of data sparsity in collaborative filtering algorithms.(4)On the basis of algorithms proposed above,this thesis develops a movie recommendation system that can record the historical ratings and the recommendation results for a user once he/she logins in.This system can automatically generate a new movie collection that is more in line with the hobby according to the rating of users on the given movie list.In addition,by collecting the rating of the current popular movie on the main interface,we can have more accurate recommendations to users.