Research On Sequential Recommendation Algorithm Based On Graph Neural Network

Sequential recommendation systems use the interaction sequences of users with items to obtain the preference of the current user sequence and accordingly predict the next possible item that the user may interact with.Dynamic graph neural networks have become a powerful approach in sequential recommendation systems,as they can effectively capture dynamic preference features from the recorded user behavior data.Most existing studies on dynamic graph neural networks for sequential recommendation rely on stacking multiple attention modules to capture dynamic collaboration signals between users and items for predicting user’s preference.However,in online platforms,the recorded user behavior data often contains noise or maliciously falsified data.As attention modules focus on the importance analysis of the entire sequence,stacking multiple attention networks can easily amplify the impact of noise data in the sequence.Additionally,new nodes will continue to appear in dynamic graphs over time,and the new node often have very few links.The scenario that predicts new links to these new nodes is called a cold start scenario in sequence recommendation.Recommendation on dynamic graph is prone to cold start,nevertheless,there is still a lack of comprehensive research on how to effectively use dynamic graph neural networks for recommendation in cold-start scenarios.In response to the aforementioned issues,the primary focus of this article is to delve into the following research topics:(1)In light of the noise amplification issue that plagues sequential recommendation algorithms based on the dynamic graph,this paper presents a sequential recommendation approach named Filter-enhanced Temporal Graph Neural Network.Based on the concept of fast discrete Fourier transform,the model designs a learnable filter that filters out noise data,and conducts feature aggregation and propagation by analyzing the correlation of the purer node representations via attention networks.In this way,the dynamic graph neural network is able to effectively capture the dynamic collaboration signals between users and items.The model is comprehensively evaluated on three Amazon review datasets and compares it with four existing recommendation system models.The experimental results demonstrate that,the model this paper proposed outperforms those sequential recommendation models,and it supporting the effectiveness of the model proposed in this paper.(2)Regarding how to effectively apply dynamic graph neural networks to cold-start recommendation scenarios.,in this paper,further combining dynamic graph neural network with meta-learning techniques,a Filtered-enhanced Temporal Graph Neural Network with meta-learning was proposed.In the model,a time interval adaptive strategy and a node adaptive strategy are designed.The meta-learning is used to adaptively adjust the parameters in the dynamic graph model,which can effectively capture the general knowledge of the nodes and the time correlation of the edges of the nodes.After meta-testing,the model can quickly adapt to new nodes,thereby reducing the impact of cold start scenarios on the efficiency of the recommendation algorithm based on the dynamic graph.The model was comprehensively evaluated on three Amazon review datasets and found that the dynamic graph model combined with meta-learning produced better recommendation results in cold-start scenarios compared to the existing four sequential recommendation models and three recommendation models used for cold start scenarios.It is proved that the combination of meta-learning and dynamic graph neural network is effective and feasible in cold start recommendation scenarios.