Application Of Personalized Recommendation Algorithm Based On User Clustering In The Shopping Guide Site

Shopping guide sites display a large number of preferential goods on web pages,which brings great convenience to consumers and business.In order to solve the sparseness problem of the score matrix,many existing recommendation systems choose to fill the score matrix with the average number or the mode,taking no advantage of personalized features of shopping guide sites,which may result in low recommendation accuracy.In addition,in order to speed up the online computing speed,many existing shopping guide sites choose to cluster users with partition-based clustering algorithms.Although the time complexities of these algorithms are relatively low,it is easy for partition-based clustering algorithms to fall into the local optimal solution and these algorithms are not sensitive to the shape of clusters.The problems mentioned above may result in the reduction of the recommendation accuracy.Aiming at the above problems,this paper proposed a personalized shopping guide recommendation algorithm based on user clustering.In the process of matrix filling and user clustering,a method of matrix filling with Naive Bayesian algorithm and an improved clustering algorithm were proposed respectively.Finally,the system performance test was carried out on the basis of realizing each function module.The main work of the paper is as follows:(1)Requirements analysis and function design of the shopping guide recommendation system.The requirements of the shopping guide recommendation system were analyzed,and the main function modules such as data collection module,behavior quantization module,personalized recommendation module and popular recommendation module were designed.The data collection module was used to collect commodity attributes and collect user behavior data with Ajax technology,which provided data basis for the personalized recommendation.The behavior quantization module summed various behaviors of the user according to the corresponding weight,which created conditions for similarity calculating and userclustering.The personalized recommendation module recommend commodities to consumers with the personalized shopping guide algorithm based on user clustering.The popular recommendation module recommend those goods with high sales to consumers.(2)The personalized shopping guide recommendation algorithm based on user clustering.This paper presented a complete personalized recommendation algorithm including steps like product category filtering,matrix filling,user clustering,and generating recommendations.In the process of matrix filling,a method of matrix filling was proposed by using Naive Bayesian algorithm aiming at the problem that default value filling methods such as the average value filling method and the mode filling method might lead to low recommendation rate.Using attributes of the commodity as the characteristic of Naive Bayesian algorithm,the problem of multi-level scoring prediction was transformed into the multiple classification,the scoring was predicted and the sparse matrix was filled.In the process of user clustering,aiming at the problem that partition-based clustering algorithms were easy to fall into the local optimal solution and were not sensitive to the shape of clusters,a bisecting K-means algorithm based on the density partition criterion was proposed.The bisecting K-means algorithm was used to alleviate the problem of local optimal solution and the DBSCAN algorithm was used to find the cluster with most sub clusters,The cluster with most sub clusters was used as the cluster to be divided further in the bisecting K-means algorithm to solve the problem that partition clustering algorithms were insensitive to the shape of clusters.(3)The realization and performance testing of the shopping guide recommendation system.We used Java and Mahout to realize those function modules and performed two experiments about the clustering algorithm and the recommendation algorithm.The experimental results showed that the bisecting K-means algorithm based on the density partition criterion could improve the clustering purity and the personalized recommendation algorithm could reduce MAE by about 12% and increase Precision and Recall by about 5% compared with the existing recommendation algorithm.