Research On Recommendation Algorithms Based On Graph Convolutional Networks

Collaborative Filtering(CF)technologies are widely used in recommender systems due to the flexibility and high performance.Traditional CF methods modeled user and item representations based on matrix factorization to obtain user's personalized interests to items,however the performance is usually limited by the sparsity of users' rating data.Recently,with the development of Graph Convolutional Networks(GCNs),graph based recommender algorithms have become a research hotspot.GCNs encode the node representations with the graph structure information by iteratively performing neighbor aggregation,which can effectively alleviate the sparsity issue of rating data.Although graph based recommendation algorithms have greatly improved the recommendation accuracy,there are still some problems to be studied.The first problem is graph nodes' attributes are incomplete,and the current attribute enhanced recommendation algorithms cannot make full use of attribute information to improve model performance.The second problem is the inconsistency in the structure of training and testing graphs.Most of the existing recommendation algorithms based on graph neural networks are straightforward and difficult to test on new users(items)without any links,that is,inductive graph recommendation problems.To this end,this thesis mainly research on follows:(1)An adaptive graph convolutional network approach for joint item recommendation and attribute inference.By constructing attributed user-item bipartite graph,we fuse the available rating and attribute information.For the incomplete attributes problem,we propose an Adaptive Graph Convolutional Network(AGCN)approach for joint item recommendation and attribute inference.The key idea of AGCN is to iteratively perform two modules: 1)the graph learning module which learn graph embedding parameters with previously learned approximated attribute values;2)the attribute update module which send the approximated updated attribute values back to the attributed graph for better graph embedding learning.Therefore,AGCN could adaptively adjust the graph embedding learning parameters by incorporating both the given attributes and the estimated attribute values,in order to provide weakly supervised information to refine the two tasks which conclude attribute inference and item recommendation.Extensive experimental results on real-world datasets clearly show the effectiveness of the proposed model.(2)An inductive graph learning framework based on transfer network.CF based recommendation algorithms hard to make predictions for new items that not appear in training.To tackle this challenge,we propose Trans GRec which an inductive graph learning framework based on transfer network.Trans GRec is composed of two parts: a graph neural network followed by an item embedding transfer network.Specifically,the graph neural network part exploits the higher-order proximity between users and segments for a better representation learning.The transfer network provides a knowledge transfer function that approximates the learned item embeddings from graph neural networks by taking each item's visual content as input,in order to tackle the new segment problem in the test phase.Extensive experimental results on a real-world video highlight dataset clearly show the effectiveness of our proposed model.Please note that,our proposed framework is generally applicable to any inductive graph based recommendation model to address the new node problem without any link structure.