首页 学者频道 学术频道 创新频道

SIGIR 2019|罗格斯大学提出一种新的策略引导路径推理方法(PGPR)

科界 06月15日

来源:学术头条

SIGIR是一个展示信息检索领域中各种新技术和新成果的重要国际论坛,SIGIR专注于信息存储,检索和传播的各个方面,包括研究战略,输出方案和系统评估。

20190615120945_973ed4.jpg

罗格斯大学发表的一篇论文《Reinforcement Knowledge Graph Reasoning forExplainable Recommendation》已被SIGIR 2019接收。文中提出了一种称为策略引导路径推理(PGPR)的方法,该方法通过在知识图中提供实际路径来结合推荐和可解释性。该论文实验结果表示,PGPR方法在NDCG, Hit Rate, Recall 和Precision方面始终优于所有数据集上的所有其他baselines。

论文题目

Reinforcement Knowledge Graph Reasoning forExplainable Recommendation

作者

Yikun Xian、Zuohui Fu、S. Muthukrishnan、Gerard de Melo、Yongfeng Zhang

会议/年份

SIGIR 2019

链接

https://arxiv.org/pdf/1906.05237.pdf

Abstract

Recent advances in personalized recommendation have sparked great interest in the exploitation of rich structured information provided by knowledge graphs. Unlike most existing approaches that only focus on leveraging knowledge graphs for more accurate recommendation, we perform explicit reasoning with knowledge for decision making so that the recommendations are generated and supported by an interpretable causal inference procedure. To this end, we propose a method called Policy-Guided Path Reasoning (PGPR), which couples recommendation and interpretability by providing actual paths in a knowledge graph. Our contributions include four aspects. We first highlight the significance of incorporating knowledge graphs into recommendation to formally define and interpret the reasoning process. Second, we propose a reinforcement learning (RL) approach featuring an innovative soft reward strategy, user-conditional action pruning and a multi-hop scoring function. Third, we design a policy-guided graph search algorithm to efficiently and effectively sample reasoning paths for recommendation. Finally, we extensively evaluate our method on several large-scale real-world benchmark datasets, obtaining favorable results compared with state-of-the-art methods.

20190615120945_994888.jpg

20190615120945_9b9ee9.jpg

20190615120945_9ee59b.jpg

20190615120946_a2f663.jpg


推理 罗格斯大学

推荐资讯