基于左归一化图卷积网络的推荐模型

Recommendation model based on left-normalized graph convolution network

图神经网络(GNN)应用于推荐领域之后取得了巨大的成功,通过堆叠多层图神经网络层聚合邻居的信息,使节点可以获取更加广阔的协同信息,可以有效解决推荐系统的数据稀疏问题。目前大部分将图卷积网络应用于推荐的工作都遵循对称归一化的设计,对称归一化加强了冷门商品对于用户嵌入的构建,但是会分配给流行度高的商品很低的权重,导致热门商品对于用户节点嵌入影响微乎其微。针对这一问题,提出一种左归一化图卷积网络模型,模型使用了更加灵活的归一化处理方式,加入衰减因子,并且设计了两种针对各图卷积层的衰减机制,相互配合,大幅提高了推荐的效果。在数据集Alibaba、Amazon-book、Yelp2018上与基准模型LightGCN、NGCF(Neural Graph Collaborative Filtering)、PinSage、Bprmf进行了对比实验,结果表明,与LightGCN相比,所提模型的召回率(recall)分别提高9.9%、20.2%、7.5%,归一化折损累计增益(NDCG)分别提高12.7%、24.1%、8.3%,验证了所提模型的有效性。

The application of Graph Neural Network(GNN)in the field of recommendation has achieved great success.By stacking multi-layer graph neural network layers to aggregate the information of neighbors,nodes can obtain broader cooperative information,which can alleviate the problem of data sparsity in the recommendation system.At present,most of the works applying graph convolutional network to recommendation follow the design of symmetric normalization.Symmetric normalization strengthens the construction of user embedding for unpopular items,but it assigns very low weight to popular items,so that popular products have little influence on user node embedding.To solve this problem,a new left-normalized graph convolution network model was proposed.The model uses a more flexible normalization processing method,adding attenuation factors,and designing two attenuation mechanisms for each graph convolution layer,which cooperate with each other to greatly improve the effect of recommendation.Comparative experiments with LightGCN,NGCF(Neural Graph Collaborative Filtering),PinSage and Bprmf were conducted on data sets Alibaba,Amazon-book,Yelp2018.Experimental results show that compared with LightGCN,the proposed model improves the index recall by 9.9%,20.2%,7.5%,and NDCG(Normalized Discounted Cummulative Gain)by 12.7%,24.1%,and 8.3%,respectively,verifying the effectiveness of the proposed model.