Research On Some Co-clustering Algorithms In Recommendation Systems

With the rapid development of science and technology,we have entered the era of big data.The network is providing us with more and more information and services.However,when we enjoy the convenience brought by the network,we also have to face a lot of junk information on the network,which is called “information overload”problem.In the face of a large number of information resources,it is currently a hot research topic on how to help users to obtain useful information quickly and effectively.Then recommendation systems come into existence.As a new way of intelligent information service,the recommendation systems can make sense of users' needs and recommend items that target users may be interested in to target users so as to realize personalized services.This greatly reduces the cost of searching information for users.In recent years,scholars have studied a large number of recommendation models and algorithms to improve the performance of personalized recommendation systems.Among them,the collaborative filtering based algorithms are widely used in many fields,such as e-commerce,social network,comment website,etc.However,with the increase of the scale of recommendation systems,collaborative filtering technology faces great challenges.Especially,the sparsity,scalability and cold start problems,which are common in the recommendation systems,become the main bottleneck of many collaborative filtering technology based algorithms.In order to alleviate these problems,this paper mainly studies the following three algorithms which are based on co-clustering to improve the performance of collaborative filtering algorithms:(1)This paper proposes a novel soft clustering algorithm,namely the soft K-indicators alternative projection(SKAP)algorithm,and establishes the SCoC recommendation method based on soft co-clustering by combining SKAP algorithm with traditional collaborative filtering recommendation algorithms.Because the use of additional information is helpful in improving the performance of the recommendation systems,this paper considers integrating the item type information into the proposed SCoC recommendation method.First,this paper utilizes the user-item rating information and the item type information to establish the user preferences information,and then combines rating information,user preferences information and the item type information to generate the co-clustering based SCoC model.(2)This paper supposes that different parts of one domain have the same rating type information and establishes the IDTL algorithm based on context information and transfer learning.First,referring to the use of the context information in existing models,this paper establishes random decision tree to partition user-item rating matrix and generates a series of sub-matrices.Then this paper selects some suitable sub-matrices and regards them as "multiple domains" to establish in-domain transfer learning.This process can generate users / items partition information and the cluster-level rating pattern shared by all sub-matrices.On the basis of this,this paper can predict the unknown ratings in user-item rating matrix.This algorithm does not need additional domain information and has stronger adaptive ability.(3)This paper considers using the orthogonal non-negative matrix tri-factorization algorithm to generate the partition information of users and items simultaneously.Then referring to the backfilling strategy in the additive co-clustering ACCAMS model,this paper establishes the hierarchical representation for users / items characteristics and further realizes the effective approximation of the large-scale user-item rating matrix.Also,considering that the use of additional information can improve the performance of the recommendation systems,this paper adds user social information and the item type information into orthogonal non-negative matrix tri-factorization process for user-item rating matrix to improve the partition precision of users / items and further improves the prediction accuracy of unknown ratings in user-item rating matrix.Above all,this paper mainly researches three co-clustering based collaborative filtering algorithms.In addition,this paper compares the proposed three algorithms with classical algorithms in recommendation systems respectively,and verifies the effectiveness of the proposed algorithms by numerical experiments.