The Research Of Personalized Recommender System Based On Spectral Clustering

With the rapid development of Web 2.0 and e-commerce, information resources have gone into the stage of exponential growth. Now recommender system is an efficient method to solve "information overload" problem. Even though collaborative filtering algorithm is the most applied technology in recommender systems, it still has faced with many problems, such as data sparse, extendibility, cold start etc.At the same time, most of the personalized recommendation systems tend to ignore some attribute information of users, such as age, gender and occupation. If obtaining user-item ratings data is difficult, this ignoring will seriously affect the recommending accuracy of personalized recommendation systems.After analysis and comparison of various common personalized recommendation algorithms and related technologies, aim to both improve the accuracy of recommendation algorithms and decrease run time of it, this thesis studied spectral clustering based personalized recommendation systems. The main contents of this paper include the following aspects.(1)This thesis introduces spectral clustering into personalized recommendation systems. A novel weighted kernel fuzzy clustering algorithm, an improved initial centroid selecting algorithm, and a revised Person correlation are combined with collaborative filtering to form two improved user-based collaborative filtering algorithms based on spectral clustering. On the dataset of MovieLens 100K, the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) of these two algorithms are at least 4% lower than the traditional collaborative filtering algorithm based on K-means clustering, and the algorithms’run time is reduced by at least half. On the dataset of MovieLens 1M, MAE and RMSE are at least 2% lower, and the algorithms’ run time is reduced by at least 80%.(2) By the use of users’attribute, the pretreatment ways of users’age, gender and occupation are presented, after obtaining the user characteristic attribute matrix, the collaborative filtering algorithm is proposed based on user characteristic property of spectral clustering.(3) Aim to tackle the over-fitting problem in Bias Singular Value Decomposition (BSVD) algorithm, using the user characteristic attribute and the user-item rating history records to combine the above proposed spectral clustering model based on the user characteristic property with BSVD, we proposed an improved recommendation algorithm which used detection of new user to solve the cold start problem. The improved algorithm was compared with BSVD algorithm on the dataset of MoveiLe-ns 100K and MovieLens 1M. The MAE and RMSE of this algorithm are at least 6% lower on the dataset of MovieLens 100K, and at least 2% lower on the dataset of MovieLens 1M. Experiments show that the algorithm not only improves the accuracy of the recommendation, but also has certain scalability.(4)Multiple experiments were designed to compare the proposed algorithms and traditional algorithms, experiment results demonstrate that the spectral clustering applied to personalized recommendation systems can greatly improve the prediction accuracy and real-time responding speed of the systems, so as to bring greater economic benefits for enterprises and businesses in the final.