Research On Sequence Recommendation Algorithm Based On Graph Neural Network

The growth of the mobile internet in recent years has led to a further increase in the type and number of online resources,which has brought convenience but also information overload for users.The problem of information overload means that it is difficult for users to find items that match their interests and preferences from the explosion of resources.Recommendation Systems(RS)are an effective means of solving the information overload problem.Previous recommendation methods use the item set that users interact with or the user-item rating matrix to learn users' preferences,which is based on the assumption that user interests are static and do not change over time.However,in practical application scenarios,the interaction behaviour between users and items is dynamic and there are temporal dependencies.Sequential RS solve problems existing in previous RS.And because it has the ability advantage of modeling temporal association in the sequence of user-item interaction,Sequential RS are able to learn user preferences effectively compared to previous RS.As a result,they are widely used in various Internet platforms.In order to better model multi-relational user-item interaction sequences,sequential RS based on Graph Neural Networks(GNN)use multi-relational sequences of historical interactions to form graph-structured data.By exploiting GNNs' ability of extracting the local items features,these models effectively improve the recommendation ability of sequence recommendation models and provide new ideas for the research of sequence recommendation algorithms.But existing sequential recommendation models based on GNN still have some limitation,specifically: firstly,previous models use short term sessions divided by long sequences to construct graphs,ignoring the temporal relationships among sessions;secondly,previous models lack modules for effective learning preference and predicting,and the individual hierarchical modules of models are poorly interpretable;thirdly,the association between preferences of multi-relational interactions is ignored;fourthly,the ability to extract global features for long sequences is lacking.For these limitations,this paper uses the Attention mechanism and Deep Neural Network(DNN)to overcome limitations of previous sequential recommendation methods focusing the sequential recommendation methods based on GNN,with main work and contributions as follows:(1)For the limitations of previous models which ignore temporal relationships among sessions and the lacks modules of effective learning preference and predicting,the GNN sequential recommendation method MGAD is proposed,which combines Attention mechanisms and DNN.Firstly,the multi-relational users' behaviours are divided into the target behaviours and the auxiliary behaviours according to the type of interaction and the relevance to the predicted behavior.The long sequence of target behaviours and the long sequence of auxiliary behaviours form together a multi-relational behaviour sequence graph.Then the Multi-relational Graph Neural Network(MGNN)is combined with a Self-Attentive(SA)mechanism and a DNN to build the model.The item representation is obtained through MGNN learn the long sequence of interactions,which avoids the problem of ignoring the temporal relationship among sessions.The preference is learned effectively by the SA mechanism.Finally,MGAD obtains more accurate prediction results through DNN.These improves the performance of recommendation.(2)For the limitations of previous models which ignore the association between multi-relational interaction preferences,based on the last work,a GNN sequential recommendation method MRHA is proposed,which combines a hybrid Attention mechanism and a DNN.Through the degree of association between the target interaction and the auxiliary interaction learns the attention scores,the Hybrid Attention(HA)mechanism can learn the user preferences from item representations with multi-relational interaction features.Thus it can improve the results of the original model on larger scale and datasets with significant interaction segmentation,based on MRGD model.(3)For the limitations of previous models whose GNN lacks the ability to extract global features for long sequences of interaction,based on the work in Chapter 4,a Local and Global Attention(LGA)mechanism-based GNN sequential recommendation method is proposed.By taking advantage of the extraction ability of the weighted attention mechanism for global temporal dependency features of long sequences,MRHA model combines the weighted attention mechanism for learning global users' target behaviour preferences.The combination of HA and GNN is used to learn local target and auxiliary behaviours preferences.The weighted attention mechanism in the LGA model improves recommendation results significantly,solving the limitation of MRHA's poor performance on small datasets.