A Lightweight GCN Recommendation Method Based On Interactive Interest Graph

In recent years,recommendation models based on Graph Convolutional Networks(GCNs)have gained popularity among researchers.By incorporating user-item graph structure information into the model,GCNs preserve latent information and effectively implement collaborative filtering concepts.However,some issues still persist.On one hand,as GCN-based recommendation models become more complex,they increasingly contradict the requirement for deep mining and utilization of basic data,as well as the research focus on reducing computational burden.On the other hand,most GCN-based recommendation models do not adequately address problems arising from loose item sequences or insufficient item interactions,as well as the majority of graph neural network-based methods suffer from oversmoothing,which greatly limits performance improvement.To tackle these issues,we propose two strategies.First,we propose a lightweight recommendation model named LG＿APIF,which employs GCN for deep information fusion.The model combines behavior memory and the Ebbinghaus forgetting curve to simulate user interest changes.Meanwhile,it leverages linear regression and other lightweight traditional methods to mine adaptive periods and other deep information for items.It also analyzes users’ current interest distributions and calculates items’ interest values to determine users’ potential interest types.By constructing a user-type-item interaction graph and incorporating a lightweight GCN technique,the model generates the final item recommendation list.Experimental results validate the effectiveness of our proposed method.Compared with 8 classical models across four datasets,LG＿APIF achieves an average improvement of 2.11% on Precision,1.01% on Recall,and 1.48% on NDCG,respectively.Second,we introduce a lightweight GCN recommendation strategy for cluster-based interest graph interaction and splitting,which incorporates user intent.By integrating user intent with time and location factors,we first construct user and item collaboration graphs to address the issue of loose item sequences.Then we center on the user-cluster interest graph and combine it with the item-cluster interest graph for intra-and inter-cluster interactions,which generates a high-density interest graph to primarily resolve the problem of insufficient item interactions.Finally,after splitting the high-density interest graph’s subgraphs twice,nodes within the subgraph are propagated with lightweight GCN to learn embedded representations and interest predictions result in alleviating the oversmoothing issue.Experiments demonstrate the effectiveness of our proposed method.Compared with 7 classical models across three datasets,our approach outperforms the competition,with average improvements of 3.8% in HR,8.1% in NDCG@5,and 4.03% in NDCG@15,respectively.In summary,within keeping model complexity manageable,our research demonstrates that extracting deep information from basic data can better construct interactive interest graphs and improve recommendation performance.Additionally,the interactive interest graphs we construct not only facilitate sufficient item interaction but also mitigate the oversmoothing problem inherent in GCNs.To this end,we propose the lightweight GCN recommendation method LG＿API which combines adaptive periods and interest factors,as well as a lightweight GCN for cluster-based interest graph interaction and splitting that incorporates user intent.Experimental results indicate that both proposed methods effectively address their respective problems and demonstrate their feasibility and effectiveness.