KASR: Knowledge-Aware Sequential Recommendation

Qingqin Wang,Yun Xiong,Yangyong Zhu,Philip S. Yu

DOI: https://doi.org/10.1007/978-3-030-60259-8_36

2020-01-01

Abstract:The goal of sequential recommendations is to capture the transitions of users’ interests. Most existing methods utilize sequential neural networks to model interaction records, mapping items into latent vectors. Although such methods do explore the transitions of items in interaction sequences, they only capture the sequence dependencies of items, neglecting the deep semantic relevance between items. Such limited information contributes less to catching the complicated sequential behaviors of users accurately. In this paper, we propose a novel model Knowledge-Aware Sequential Recommendation (KASR), which captures sequence dependencies and semantic relevance of items simultaneously in an end-to-end manner. Specifically, we first convert the interaction records into a knowledge-transfer interaction sequence, which reflects the fine-grained transitions of users’ interests. Next, we further recursively aggregate information in the knowledge graph based on a specific relation attention network, to explicitly capture the high-order relevance between items. A knowledge-aware GRU is later introduced to explore the sequential and semantic relevance between items automatically. We have conducted extensive experiments on three real datasets, and the results demonstrate that our method outperforms the state-of-the-art models.