Knowledge Graph Extrapolation Network with Transductive Learning for Recommendation

Abstract

A knowledge graph is introduced into the personalized recommendation algorithm due to its strong ability to express structural information and exploit side information. However, there is a long tail phenomenon and data sparsity in real knowledge graphs, and most items are related to only a few triples. This results in a significant reduction in the amount of data available for training, and makes it difficult to make accurate recommendations. Motivated by these limitations, the Knowledge Graph Extrapolation Network with Transductive Learning for Recommendation (KGET) is proposed to improve recommendation quality. To be specific, the method first learns the embedding of users and items by knowledge propagation combined with collaborative signal to obtain high-order structural information, and the attention mechanism is used to distinguish the contributions of different neighbor nodes in propagation. In order to better solve with data sparsity and long tail phenomenon, transductive learning is designed to model links between unknown items to enrich feature representation to further extrapolate the knowledge graph. We conduct experiments with two datasets about music and books, the experiment results reveal that our proposed method outperforms state-of-the-art recommendation methods. KGET also achieves strong and stable performance in sparse data scenarios where items have merely a few triples.