Research On Sequential Recommendation Algorithm Based On Information-enhanced Graph Neural Network

Sequential recommendation algorithms predict the next item that the user may be interested in according to the user’s historical interaction records,regard the user behavior as a dynamic sequence,and believe that there is a sequence dependency between the interactive items,so as to capture the user’s recent preference.Due to the advantages of graph neural network in representation learning and the interaction order between users and items can be expressed as sequence graph structure,sequential recommendation algorithms based on graph neural network attract extensive attention.However,most of the existing sequential recommendation algorithms based on graph neural network only focus on the structural information of user’s interaction with items,and the learning of sequence preference only involves the sequence of interactive items,which lacks the content information of the items themselves and the effective utilization of user’s information,and does not mine the deeper semantic relationship between items.In addition,items in the sequence can be divided into categories in real life,and the interaction time interval of items also reflects the correlation between items.The effective utilization of item category information and time interval information will help to improve the performance of sequential recommendation algorithm.To solve the above problems,this thesis proposes a sequential recommendation algorithm based on information-enhanced graph neural network,which integrates side information such as knowledge graph,category and time into graph neural network to model,and it is helpful to improve the accuracy of sequence recommendation.The main works of this thesis are as follows:(1)This thesis proposes a sequential recommendation model based on knowledge-aware graph neural network.Aiming at the lack of item’s content information and user’s information,this thesis proposes to introduce knowledge graph,which is combined with user’s interaction data to construct collaborative knowledge graph.The model uses the graph attention network to learn item’s semantic relevant auxiliary information and user’s relevant auxiliary information based on collaborative knowledge graph.Then,the interactive sequence is constructed into a sequence graph,the gated graph neural network is used to fuse the user auxiliary information to learn the sequence preference,and combined with the items semantic auxiliary information to predict,which effectively improves the accuracy of recommendation.(2)This thesis proposes a sequential recommendation model based on multi feature aggregation,which learns features from two dimensions of category and time to optimize.For category information,on the one hand,the category sequence is obtained from the item sequence to construct the category sequence graph and learn the categories interest transfer preference.On the other hand,the item sequence is divided into multiple sub sequences according to the categories,and each sub sequence is modeled separately to obtain more fine-grained sequence preference under the same category context.For time information,this method obtains the time intervals between the items in the sequence and constructs the time-aware sequence graph,which measures the correlation between items from the time dimension,and designs the time gate to model time interval based on the gated graph neural network.A series of experiments are carried out on three public datasets: Amazon-Book,Last-FM and Yelp2018.The metrics are hit rate(HR@K)and normalized discounted cumulative gain(NDCG@K).Compared with the optimal method in the baselines,the metrics of the algorithm proposed in this thesis have an increase of at least 0.15 percentage point on Last-FM dataset with stronger sequence,at least 5 percentage points on Yelp2018 dataset with weaker sequence,and at least 16 percentage points on Amazon-Book dataset with relatively weakest sequence.The experimental results show the effectiveness of the method proposed in this thesis,and the improvement effect of this method is more obvious in the fields with weak sequence.