Conversation Recommender System Based on Knowledge Graph and Time-series Feature

Abstract

The conversation recommender system aims to recommend high-quality items to users through interactive conversations, which requires to be seamlessly integrated between the recommendation module and the dialog module. The knowledge graph has improved the accuracy of the dialogue recommendation system to a certain extent. However, there are still some defects that make it easy to generate more general and popular responses. For overcoming these shortcomings, this paper proposes a novel framework based on knowledge graphs and time-series features called KGTF. By serializing and modelling the dialogue content, the KGTF is able to learn the feature relationship between users and items to make more accurate recommendation. Besides, the location encoding is introduced to increase the diversity of generated responses. Experiments conducted on widely adopted benchmarks show that the proposed KGTF framework is superior to the latest KGSF method.