Abstract:

Knowledge graph can effectively alleviate problems of data sparse and cold start in traditional collaborative filtering. Therefore, it has become an important exploration direction to integrate knowledge graph into the recommender system. However, most of the methods divide the network structure of knowledge graph into separate paths or only use the first-order neighbor information, which make it impossible to establish the high-order connectivity on the whole graph. To solve the problem, this paper proposes a KG-BGAT model which combines knowledge graph and graph attention network, and designs a bilinear collector. The bilinear collector can obtain the feature interaction message between nodes during the stage of information collection, and enrich the representation of nodes. The graph attention network propagates each node representation along the graph through the recursive embedding propagation algorithm, which can capture the high order connectivity in the graph. Top-K recommendation experiments are tested on the Movielens-1M dataset. When the length of the recommendation list is 20, the accuracy rate, recall rate and normalized discounted cumulative gain are 29.4%, 24.9% and 67.4% respectively, which exceeds the current mainstream recommendation algorithms such as CKE, RippleNet and KGCN. Experiments show that the proposed method can improve the accuracy of the recommended results.

Key words: recommendation system, knowledge graph, feature interaction, graph attention network

摘要：

知识图谱可有效缓解传统协同过滤中的数据稀疏和冷启动问题，因此，近年来在推荐系统中融入知识图谱的方法成为重要的探索方向。然而现有的方法大多将知识图谱的网络结构划分为单独路径或仅利用了一阶邻居信息，造成无法建立整个图上的高阶连通性问题。为解决该问题，提出融合知识图谱和图注意力网络的KG-BGAT模型，并设计了双线性采集器。双线性采集器能够在信息采集阶段获取节点间的特征交互信息，丰富节点表示；图注意力网络通过递归嵌入传播算法将各个节点表示沿图进行传播，能够捕获图中的高阶连通性。在MovieLens-1M数据集上进行了Top-K推荐实验，在推荐列表长度为20时，精确率、召回率和归一化折损累计增益分别为29.4%、24.9%、67.4%，超过了目前主流的CKE、RippleNet、KGCN等融合知识图谱的推荐算法。实验证明提出的方法能够有效提高推荐结果的准确性。

关键词: 推荐系统, 知识图谱, 特征交互, 图注意力网络