Lightweight Graph Convolutional Collaborative Filtering Recommendation Approach Incorporating Social Relationships

Recommendation systems are a key technology for solving the information overload problem.The focus of recommendation is on predicting user preferences and broadening user horizons.Existing recommendation methods mainly use explicit feedback to directly reflect user preferences or implicit feedback to indirectly reflect user preferences to make recommendations.However,recommendation models with explicit feedback generally cannot take into account differences in users’ tolerance for scoring mechanisms,which can,to a certain extent,cause data errors and thus affect recommendation results.Researchers have therefore started to make use of implicit feedback information,but when using implicit feedback,they often ignore implicit negative feedback data that users do not view or click,which is denser and can indirectly tap into users’ hidden preferences,with the disadvantage that it is not easily accessible and more likely to contain noisy data.In recent years,graph convolutional networks have been rapidly developed for their powerful modelling capabilities,and can effectively aggregate first-order neighbourhood information in user-item interaction data through representation learning.However,most of the current recommendation models directly inherit the complex design of graphical convolutional networks(e.g.feature transformations,non-linear activation,etc.),with a complex model structure,a large number of training parameters,and mostly a lack of modelling of higher-order interaction features between users and items.To deal with such problems,this paper proposes a lightweight graph convolutional collaborative filtering recommendation model incorporating social relationships(called for F-Light GCCF).The proposed model maps user,item,and friend information to a low-dimensional dense vector space through graph embedding technique in the embedding layer,which alleviates the negative impact of data sparsity on the recommendation results of the model.The topology of user social relationship graph(i.e.,user-item-friend higher-order connectivity graph)is learned by stacking three graph convolution layers in the graph convolution layer to generate a series of indirect feedbacks from implicit negative feedbacks,and the implicit negative feedbacks are captured indirectly by analyzing user behavior and friend closeness,which expands the training data and improves the utilization of implicit negative feedbacks.The fusion graph attention network to measure the contribution value of neighbors and assign weights to them adaptively and dynamically can filter noisy neighbors to make them robust,which helps to select the most relevant information instead of all available information.At the same time,the complex design of the graph convolution layer is simplified,and the design of feature transformation and nonlinear activation in standard graph convolutional neural networks is discarded,which simplifies the training difficulty of the model and improves the model effect.Besides,the model prediction layer is designed with a hierarchical aggregation mechanism to weight and aggregate the multiple embedding vectors learned in the graph convolution layer,and an attention mechanism is introduced to automatically learn the importance of the embedding vectors in the l-th layer.Last,the inner product operation is used to obtain the association score between user-items.By conducting extensive experiments on Gowalla dataset and Yelp2018 dataset.The experimental results show that the recommendation effect of the proposed method in this paper is better than the current recommendation algorithms.And then,The efficiency and effectiveness of our approach are demonstrated by experimental result..Meanwhile,the model in this paper significantly outperforms other baseline algorithms in Recall@k,Precision@k,and NDCG@k metrics,which proves the reasonableness of the model design.This paper has 41 pictures,6 table,and 99 references.