Abstract:

With increasing in scales of social networks, modeling complex graphs come to be a challenge for cold start recommendations. The conflicts involving in these complex information will directly affect the recommendation results. So the paper proposes a Graph Neural Networks recommendation algorithm based on Three-Way Decision theory, named as TWD-GNN. Three-way decision is introduced to divide the entire data set into tri-partitions as positive, boundary, and negative domains. And solving local conflicts can benefit from the progress, which the auxiliary information from multi-sources is integrated with the boundary domain. The graph neural network method is available for effectively mining information, reconstructing the network and predicting the scores for recommendations. The experimental results show that the proposed algorithm can reflect user preferences more accurately than the traditional collaborative filtering recommendation algorithms, and and improve the recommendation quality.

Key words: recommendation system, information conflict, graph neural network, three-way decision

摘要：

随着网络规模的不断扩大，对复杂图结构进行建模是推荐任务面临的一大挑战。这些复杂信息之间易存在冲突，往往会直接影响推荐结果。为此，提出了基于三支决策的图神经网络推荐方法（TWD-GNN）。引入三支决策理论对数据集进行划分为正域、边界域和负域，在边界域的基础上融合辅助信息，梳理局部冲突。采用图神经网络挖掘有效信息，重构网络并预测评分。实验结果表明，TWD-GNN算法较传统的协同过滤推荐算法能更准确地反应用户偏好，提高了推荐质量。

关键词: 推荐系统, 信息冲突, 图神经网络, 三支决策