融合知识图谱的KGCN推荐算法优化研究
Research of KGCN Recommendation Algorithm Optimization Based on Knowledge Graph
符玉飞 1周莲英 1丁腊春2
作者信息
- 1. 江苏大学计算机科学与通信工程学院 镇江 212013
- 2. 江苏省镇江市第四人民医院 镇江 212001
- 折叠
摘要
KGCN通过图卷积挖掘知识图谱中的实体邻域特征,但是在构建实体邻域时所采用的均匀采样方法忽略了各邻居节点自身性质的差异.针对此问题,论文提出了一种重要邻居采样的非均匀采样方法,依据节点度中心性,实现图网络实体邻域特征挖掘中邻居节点的优选和淘汰,以此提高样本节点的选取质量.此外,在图卷积网络输入层之前,增加降噪自编码器,降低误点击样本造成的噪声干扰,以此增加训练过程的稳定性、防止过拟合.最后,论文给出了推荐算法改进模型,该模型利用神经网络的层级传播机制,可迭代聚合图网络中更广范围内节点信息,优化推荐效果.论文分别在电影和书籍两个数据集上对改进的推荐算法进行了对比性实验,实验结果表明改进的算法提高了推荐的准确性和鲁棒性.
Abstract
KGCN uses graph convolution to mine the entity neighborhood features in the knowledge graph,but the uniform sampling method used in constructing the entity neighborhood ignores the differences of the properties of each neighbor node.To solve this problem,this paper proposes a non-uniform sampling method of important neighbor sampling,according to the centrality of node degree,the optimization and elimination of neighbor nodes in graph network entity neighborhood feature mining are real-ized,so as to improve the selection quality of sample nodes.In addition,before the input layer of the graph convolution network,a denoising auto encoder is added to reduce the noise interference caused by clicking on the sample by mistake,so as to increase the stability of the training process and prevent over fitting.Finally,this paper presents an improved recommendation algorithm model,which uses the hierarchical propagation mechanism of neural network to iteratively aggregate the node information in a wider range in the graph network and optimize the recommendation effect.This paper makes comparative experiments on the improved recom-mendation algorithm on two data sets of movies and books,the experimental results show that the improved algorithm improves the accuracy and robustness of recommendation.
关键词
KGCN/知识图谱/度中心性/降噪自编码器Key words
KGCN/knowledge graph/degree centrality/denoising auto encoder引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据