融合负采样和消息传递的GCN推荐算法

GCN Recommendation Algorithm Combining Negative Sampling and Message Passing

近年来,图卷积神经网络(Graph Convolutional Networks,GCN)在推荐领域有广泛的应用,LightGCN通过对传统GCN的简化,省略特征变换和非线性激活的过程,对GCN的研究提供了新思路。为了解决推荐算法负采样问题和消息传递对GCN收敛的影响,提出了SNGCN模型,改变了直接从数据中采样原始负样本的采样策略,利用正例混合和样本混合两个步骤合成硬负样本;其次,SNGCN利用约束损失逼近多层图卷积的极限。最终该模型在4个公开基准数据集上得出的实验结果表明,其Recall和NDCG两项指标均比对比的推荐算法有所提升。

In recent years,Graph Convolutional Neural Networks(GCN)has been widely used in the recommendation field.LightGCN provides new ideas for the research of GCN by simplifying the traditional GCN and omitting the process of feature transformation and nonlinear activation.In order to solve the negative sampling problem of recommendation algorithm and the influence of message passing on GCN convergence,SNGCN model is proposed,which changes the sampling strategy of sampling original negative samples directly from data and synthesizes hard negative samples using two steps of positive example mixing and sample mixing.Secondly,SNGCN uses constrained loss to approximate the limit of multi-layer graph convolution.The final experimental results derived from this model on four publicly available benchmark datasets show that both its Recall and NDCG metrics are improved over the compared recommendation algorithms.