Scholarlink

<p>In recommendation systems, mining the implicit relationship between users and items is the key to achieving accurate recommendations. Graph structures can represent users, items, and even the interactive relationship between the two entities. Graph neural networks (GNNs) have inherent advantages in processing graph-structured data, particularly in capturing the connection relationships between graph nodes and learning representations of graph data. However, calculating the similarity between graph structures has been a persistent challenge in improving the performance of graph neural network-based recommendation systems. We propose a novel recommendation model based on GNNs called KerGNNT, where Transformers are employed to optimize the representation learning method for the graph. We extract node-centered subgraph and use graph kernels to compute kernel values of each node in the input graph to update the subgraph representation. The graph kernel allows for a more precise comparison of the similarity between graph filters and input subgraphs. Furthermore, the proposed model uses Transformer to optimize the graph representation learning. In order to capture the interactive information between nodes, the output of the GNN is connected to the Transformer, dual-encoding the position and structure of subgraphs to further enhance graph representation. We conduct a series of experiments on four different datasets of different scales and categories (COLLAB, IMDB-BINARY, Reddit, and ZINC), and the experimental results demonstrate that our proposed model performs better in terms of mean absolute error (MAE) and mean squared error (MSE), indicating an advantage in the accuracy of recommendations. KerGNNT also exhibits good interpretability. Visualization through GNN filters helps us understand the features that GNN focuses on. The experimental results also verify that the features extracted by our proposed model are positively correlated with improving the recommendation performance. Compared with baseline models, our proposed model reduces the MAE by at least 51.2% and the MSE by at least 51.6%, which indicates the improvement of recommendation performance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

KerGNNT: an interpretable graph neural network recommendation model

  • 期刊论文

  • 发表日期：

    2025年3月10日

查看全文
  • Shengbin Liang,
  • Jinfeng Ma,
  • Fuqi Sun,
  • Tingting Chen,
  • Xixi Lu,
  • Shuanglong Ren

摘要

In recommendation systems, mining the implicit relationship between users and items is the key to achieving accurate recommendations. Graph structures can represent users, items, and even the interactive relationship between the two entities. Graph neural networks (GNNs) have inherent advantages in processing graph-structured data, particularly in capturing the connection relationships between graph nodes and learning representations of graph data. However, calculating the similarity between graph structures has been a persistent challenge in improving the performance of graph neural network-based recommendation systems. We propose a novel recommendation model based on GNNs called KerGNNT, where Transformers are employed to optimize the representation learning method for the graph. We extract node-centered subgraph and use graph kernels to compute kernel values of each node in the input graph to update the subgraph representation. The graph kernel allows for a more precise comparison of the similarity between graph filters and input subgraphs. Furthermore, the proposed model uses Transformer to optimize the graph representation learning. In order to capture the interactive information between nodes, the output of the GNN is connected to the Transformer, dual-encoding the position and structure of subgraphs to further enhance graph representation. We conduct a series of experiments on four different datasets of different scales and categories (COLLAB, IMDB-BINARY, Reddit, and ZINC), and the experimental results demonstrate that our proposed model performs better in terms of mean absolute error (MAE) and mean squared error (MSE), indicating an advantage in the accuracy of recommendations. KerGNNT also exhibits good interpretability. Visualization through GNN filters helps us understand the features that GNN focuses on. The experimental results also verify that the features extracted by our proposed model are positively correlated with improving the recommendation performance. Compared with baseline models, our proposed model reduces the MAE by at least 51.2% and the MSE by at least 51.6%, which indicates the improvement of recommendation performance.