Research On Cross-Domain Recommendation Based On Transfer Learning

In the current information age,the problem of information overload becomes more and more serious,which makes it hard for users to quickly and effectively get the information that they need.By providing users with personalized recommendation service,recommender systems can effectively solve the problem of information overload.Collaborative filtering is one of the most successful and most widely used recommendation technologies in recommender systems at present,but it relies on the historical rating data of users,the problem of data sparseness is its major bottleneck.Transfer learning and cross-domain recommendation with the help of auxiliary data associated across domains with target data is an effective way to solve this problem.For some deficiencies of the traditional cross-domain recommendation models,we propose three new cross-domain recommendation models based on transfer learning and a new algorithm for influence maximization in social networks,and then combine this algorithm for influence maximization with cross-domain recommendation models proposed in this paper to form two-stage cross-domain recommendation models.The main contributions of this paper are as follows:(1)For the problem that the rating data in target domain and that in auxiliary domain are heterogeneous,we propose a cross-domain collaborative recommendation model based on transfer learning of heterogeneous feedbacks of users.To transfer the information of heterogeneous feedbacks of users from the auxiliary domain to the target domain,firstly we conduct normalization preprocessing and collective factorizations on rating matrices of the target data and its auxiliary data,and get the initial latent factors of users and items,when we consider both common and domain-specific rating patterns so that the negative transfer is reduced effectively.Based on the initial latent factors,then we calculate the similarity by the Pearson correlation coefficient and construct similarity graphs.Finally we predict the missing ratings in the target data by transfer learning based on the graph regularized weighted nonnegative matrix tri-factorization.Experiments show that this model can improve the transfer learning of heterogeneous feedbacks and increase the prediction accuracy for missing ratings effectively,which is more significant when the data is sparser.(2)For the problems that traditional models generate too more negative transfer or insufficient positive transfer,we propose a cross-domain recommendation model based on triple-bridge transfer learning.Firstly we extract latent factor and shared rating pattern of users as well as latent factor and shared rated pattern of items by collective factorizations on rating matrices,while considering domain-specific patterns and imposing similarity constraints on latent factors;Then we construct adjacency graphs utilizing clustering information contained in latent factors;Finally we predict the missing ratings jointly by user-side and item-side triple-bridge transfer learning based on shared pattern,latent factors and adjacency graphs.Our model reduces the negative transfer efficiently and increases positive transfer significantly,and experiments show that the prediction accuracy of this model for missing ratings outperforms several state-of-the-art recommendation models.(3)For the problem that traditional models conduct transfer learning based on non-adaptive user clustering and the ranking prediction is not accurate enough in them,we propose a cross-domain ranking recommendation model based on adaptive learning of the similarity between users across domains.Firstly we calculate the similarity between users and the similarity between user clusters at the same time based on the adaptive clustering on users.Then we weigh the latent factors of users by the similarity between users across domains to impose the cross-domain similarity constraint,and transfer the feature information between users across domains.Finally we establish a cross-domain ranking recommendation model which combines the rating prediction and list prediction.Experiments show that the ranking prediction's accuracy of this model outperforms several state-of-the-art models,and the sparseness problem of target rating data in ranking recommendation can be solved effectively by this model.(4)For the problems that traditional algorithms for influence maximization in social networks don't get widely spreading range of influence or have high time complexity,we propose a algorithm for influence maximization in social networks based on 3-layer local centrality,and build cross-domain recommendation models fused by influence maximization to solve the promotion problem of items with sparse rating data in the target domain systematically.We propose a new algorithm for influence maximization in social networks based on 3-layer local centrality.Firstly we define the 3-layer local centrality for computing the potential influence of nodes.Based on the linear threshold model,then we select some seed nodes by the heuristic method in which the node with maximal potential influence is selected for activation at each time.We choose another seed nodes later by the greedy method in which the node with the largest influence increment is chosen for activation at each time.The experimental results show that compared with some advanced algorithms,our proposed algorithm has wider activation range and lower time complexity.This algorithm can be combined with our proposed cross-domain recommendation models in the front to form two-stage cross-domain recommendation models.We can choose a group of most influential users by the proposed algorithm for influence maximization and then recommend the favorite items for them by the proposed cross-domain recommendation models in the front,so as to realize the maximization of the promotion benefit of items when the rating data in the target domain is sparse.