Research And Application Of Recommendation Algorithm Based On Three-way Clustering

The information on the Internet is increasing rapidly because of the development of information technology.This phenomenon of decreasing information acquisition efficiency and low utilization rate as the scale of information increases is called information overload.People have to waste more time filtering information by themselves.Among the methods to deal with the challenge of information overload,the collaborative filtering recommendation system has been more and more researched and applied because this method can quickly find the items of interest according to the user's hobbies.Collaborative filtering algorithms still have scalability problems in the face of increasing data.Although the introduction of clustering algorithms can alleviate the scalability challenges to a certain extent at the cost of loss of accuracy.It is because that the clustering results will not only reduce the user's nearest neighbor space but also affect the quality of nearest neighbors.And the cluster results become vulnerable to interference from some fringe users because the traditional clustering algorithm uses a single clustering rule to group all users without distinction.Meanwhile,the similarity computing problem between two users without common scoring items increases the impact of data sparsity.The main research results are as follows:1.A user clustering method for sparse data in the recommended system is proposed.Considering the data sparsity in the recommendation system,the theory of three-way decision is introduced into the clustering algorithm,which divides the users into two categories of clustering rules before clustering.The method is a soft clustering method that allows users to belong to different clusters.The proposed method can be described by two parts: first,the users set is classified as the core users and the fringes users,then the fringes users will be divide into the clusters after core users clustering;since every cluster can provide a prediction score,it's necessary to aggregate those scores to produces the final prediction score.2.When only using rating data to calculate users' similarity,it will be affected by the lack of common ratings.Therefore,this thesis proposed a hybrid recommendation algorithm integrating user preference information.First,the user movie rating matrix and the movie attribute matrix are used to obtain a dense low-dimensional matrix,which can generate similarity scores of user preferences.Then,a new similarity score is obtained by fusing these two similarity scores.Finally,we use the fusion similarity score to generate new recommendations.3.We launched experiments to examine the feasibility and validity of the algorithm,besides,a prototype system based on the Flask framework is designed and implemented in this thesis.This system is composed of key algorithm and front-end interface,the recommendation results can be displayed through the front-end interface,which proves the practicability of the algorithm.