Research On Recommendation Techniques Based On Clustering

With the growth of information on the Internet and the changes in user interactive mode,the recommender system,for which the performance requirements is also increasing,is now playing an important role in the modern Internet.As the number of users and items in the system increases,the clustering algorithm has been introduced into the recommender system so as to enhance the performance and lower the computation cost.This paper mainly studies the recommendation technology based on clustering.By clustering users and items in matrix factorization model,our work promotes the performance of recommender accuracy.The main work is shown as follows.1.This paper analysis the difference biases a user have on different kinds of items.We proposed a cluster bias matrix factorization(CBMF)model,in which a user have different bias parameters on different item clusters,so as to better extract the user's feature on different items and enhance performance.In real applications,the system performance will gradually decrease due to the continuously generated new ratings.As traditional MF model is stationary,the only way to discover the patterns of the new data is to rebuild the whole model.To solve this problem,we proposed an ICBMF model which updates the bias parameters incrementally from the newly arrived data.The experiment shows that the ICBMF model keeps the recommendation accuracy with low computation cost.2.This paper also proposed a cluster based local matrix factorization(CBLMF)model.In our model users are divided into different communities and the rating matrix are accordingly spitted into sub-matrices.Then MF model is applied on each submatrix separately.The experiment result shows that using the local model,the general performance become worse,except some clusters.In order to solve this problem,a dynamic adjust process was introduced to form a DACBLMF model.During the training,a user will change from one cluster to another if the training error decreases.The result shows that this dynamic model performs better than CBLMF in general.At last,we proposed a mixed model to combine the local and global model together.The mixed model will select the most suitable model for each user.As the experiment result shows,the mixed model outperforms both the DACBLMF model and the traditional global MF model.