Collaborative Filtering Recommendation Based On User's Preference

Recommender system automatically provides users with products or service which satisfy users' needs or taste and has become an important technology for solving the problem of information explosion.As one of the most successful recommendation techniques,recommendation technology based on collaborative filtering has received continuous attention from industry and academia due to its simple implementation and universality.However,as data sparsity and cold-start problems exist and ratings can only carry limited information,the recommendation accuracy of collaborative filtering based recommendation technology still needs to be improved.Therefore,domestic and foreign researchers have proposed many methods in order to improve its performance.Among them,how to accurately analyze users' preferences is the key for improving collaborative filtering recommendation technology.Aiming at the defects and problems existing in the collaborative filtering based personalized recommendation technology,this paper proposes methods to improve the recommendation accuracy from the following aspects after the in-depth and systematic research:(1)The existing collaborative filtering based personalized recommendation technology can only use users' ratings to their shared items and is not able to use all the scoring records,which limits the accuracy of the recommendation system,and becomes more challenging in the case of sparse data.To solve this problem,a method to decompose users' scores on the items into users' ratings to the internal information of all the items is proposed.And using this method,all the scoring records can be utilized to extract user's preferences on items' internal information.By dynamically mixing the similarity of users' ratings to their commonly-rated items with the similarity of users' ratings to items' internal information,a new recommendation method based on users' preference for items' internal information is proposed to improve the collaborative filtering technology based on users' ratings to their commonly-rated items.In addition,for the items without available internal information or for which the internal information is difficult to obtain,a method to infer the weighted internal information from users' generated tags to the items is proposed and then the weighted internal information is applied to the proposed recommendation method using all the scores.The experiments show that the collaborative filtering recommendation based on all the ratings overpowers the collaborative filtering recommendation based on user's scores on their commonly-rated items.(2)The users' ratings only reflect the users' overall quantitative preference for the items,but cannot express the details of the user's preference,which can be extracted from users' review texts to the rated items.Therefore,using the new semantic similarity calculation method in th natural language processing technology to analyze the similarity of users' review texts as the similarity of users' preferences in users' reviews,this paper designs a method to combine the semantic similarity with Pearson correlation coefficient and a heuristic similarity,which are then applied to the recommendation,respectively.Experimental results show that the proposed collaborative filtering recommendation method taking into account of users' review texts and ratings can improve the accuracy of the collaborative filtering recommendation technology.(3)In order to solve the problem that the existing time-related recommendation techniques have not considered the timing sequence of users' ratings,a method to convert the users' scores into time series data is proposed according to the chronological order.Based on the similarity calculation method of time series data for classification problems,a new similarity calculation method is proposed based on the time sequences of users' ratings.The hybrid similarity mixing the similarity based on the interest sequences with the existing similarity based on users' ratings is designed and further utilized into the collaborative filtering recommendation technology.The experiments on the real data sets verify the effectiveness of the interest sequences for improving the recommendation accuracy.(4)The method to fast calculate the similarity of users' temporal preferences is studied,and a combined recommendation method based on users' temporal preferences and items' relationship is proposed.Aiming at solving the problem that the existing similarity calculation methods for temporal trajectories have highly computational complexity and cannot support incremental calculation,this paper proposes a simhash-based method to convert users' trajectories into the trajectory fingerprints.In this method,the inverse trajectory frequency is used as the weight of the trajectory points in the trajectories and the similarity based on trajectory fingerprints is used as users' similarity.In addition,in order to reduce the influence of the popular trajectory points on the calculation of trajectory points' similarities,a dynamic penalty function according to the trajectory points' popularity is introduced into the calculation.Furthermore,a new recommendation method using the similarity of the trajectory fingerprint and the similarity of the trajectory points is proposed to improve the accuracy of the recommendation technology.The comparisons of the proposed method with the existing methods show that the method is effective.(5)Traditional recommender systems assume that users' preferences are independent without considering the influence between users.In order to remedy the drawback,the influence between users' preferences is studied.By integrating the temporal influence between users' ratings and differences of users' ratings into probabilistic matrix factorization,a novel collaborative filtering recommendation method based on the influence of users' preferences is proposed.The experiments show that the influence of users' preferences can be used to improve the collaborative filtering recommendation.