Individual&Group Recommendation Based On Graph Embedding

The recommendation system has been a very hot research direction in the past decade.The fast development of representation learning in recent years also greatly promotes the development of recommendation system.In this thesis,we mainly research on applying graph embedding,a novel representation learning technique,to two common recommenda-tion problems:one is item or friend recommendation for individual users,the other is item recommendation for group users.Traditional individual recommendation mainly relies on collaborative filtering algo-rithm,which has to face three challenging problems inevitably:(1)data sparsity problem;(2)recommendation efficiency problem;(3)complicated decision-making process of users.That is,in the social network,users may consider not only their own interests,but also their friends' choices in their decision-making process for choosing items.Also,users may con-sider the user-item interaction in choosing their friends.Given all the challenges above,this thesis proposes two novel individual recommendation algorithms.For challenging problems(1)and(2),we propose an efficient adaptive hybrid collaborative filtering algorithm.To al-leviate the problem of data sparsity,the algorithm utilizes the similarity calculation method for tensor of item content to obtain the content-based recommendation results as auxiliary information.At the same time,in order to improve the efficiency,this algorithm adopts a superposition algorithm of short path on the graph.Moreover,the adaptive hybrid method is used to integrate the recommendation results of the two methods above to improve the recommendation performance.In addition,for challenging problems(1)and(3),we pro-pose a unified graph embedding-based algorithm for joint friend and item recommendation.The algorithm first constructs two kinds of implicit user relationship graphs by utilizing the second order similarity of users in explicit user-item interaction graph and user-user social graph separately,which aims at alleviating the data sparsity problem.Moreover,the algo-rithm proposes a unified graph embedding-based model to simulate the complex decision-making process of users.Finally,the feature vectors of users and items are obtained by training model and then be used in recommendation system.For group(especially occasional group)recommendation,there are mainly two chal-lenging problems:(1)aggregation strategy choice;(2)cold start and data sparsity issues,an occasional group can be composed at any time without any historical records.To deal with the two problems above,this thesis first proposes a reasonable hypothesis that members in the group have different influences and will change in different groups.Then,this thesis proposes a graph embedding-based item recommendation solution for occasional groups.The solution first makes use of the bipartite graph embedding technology to fully learn the users' and items' feature vectors in user-item interaction graph.Secondly,it utilizes social network structure to learn the user's social influence,and then employs attention mechanism to calculate the user's influence in each group,accordingly the users' personal feature vec-tors are merged to obtain a group feature vector.These practices can solve the above two problems effectively.We evaluate the effectiveness of our algorithms on several real-world datasets,and the experimental results show that our proposed recormendation systems for individuals or groups can achieve higher accuracy than the existing state-of-art recommendation algo-rithms.