Graph Neural Network Based Long- And Short-Term Sequential Recommendation

In recent years,with the advent of the era of big data,recommender systems,as an effective solution to the problem of information overload,can help users quickly identify personal interests from massive amounts of data,which has become a hot research topic.In particular,sequential recommendation is of important research significance.It models user interests based on users’ historical interaction data so as to provide users with a recommendation list that matches their preferences.In practice,a user’s interaction sequence can be divided into a long-term interaction sequence and a short-term sequence of the most recent interactions according to the interaction timestamps.Sequential recommender systems have the following problems:(1)it still have shortcomings in capturing users’ long-term interests and short-term preferences and do not consider the impact of noise in historical interactions on user preference modeling.(2)users’ short-term preferences usually change rapidly over time,so it is difficult to capture their current preferences from historical sequences.In response to the above problems,this thesis proposes a long-term and short-term sequential recommendation model based on graph neural networks,which is called GLS-SR.The main research content includes:(1)According to interaction timestamps,a user interaction sequence is divided into a long-term historical interaction sequence and a short-term sequence consisting of the most recent interactions.We construct a compact item-item graph from long-term historical interaction sequences based on metric learning,and use user embeddings to distinguish noisy sequences.Next,the item information in the graph is updated through a graph neural network,and then the graph pooling technology is used to aggregate the user’s interest in the graph while maintaining the original structure of the graph.Finally,an aggregation function is used to obtain the user’s long-term interest,leading to improved recommendation rationality.(2)Dissertation use a sliding window strategy to divide the short-term sequence of the user’s recent interactions,use a two-layer gated network to capture the item features related to the current preferences,and finally use an aggregation function to obtain the user’s short-term preference and improve the accuracy of recommendations.(3)A gated network is used to adaptively control the contribution of users’ long-term interests and short-term preferences,and a bilinear model is used on the short-term sequence to capture the co-occurrence patterns of items in the user interaction sequence.Finally,Dissertation conduct a comprehensive experimental evaluation on three real-world user interaction behavior data sets,including Amazons,MovieLens and GoodReads.The experimental results show that the proposed GLS-SR model outperforms representative competing models in terms of Recall and NDCG.