Faithfully Explainable Recommendation via Neural Logic Reasoning

[jchanxtarov]

Basic Information

• Authors: Yaxin Zhu, Yikun Xian, Zuohui Fu, Gerard de Melo, Yongfeng Zhang
• Date: June, 2021
• Published By: Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies

Link

https://aclanthology.org/2021.naacl-main.245.pdf

Overview

• The first work that models and evaluates faithfully explainable recommendation under the framework of KG reasoning. -> logic reasoning for explainable recommendation (LOGER) = an interpretable neural logic model for logical reasoning & graph encoder to learn KG representations, trained by EM algorithm
• ex: PGPR cannot consider personalized user behavior and lack of considering the faithfulness of explanations.

Others

• Why "faithfully"? -> Because they measured faithfulness: sample around 1,000 paths connecting item nodes for each 50 users, and sample 20 paths in test (prediction) phase. And compared the rule distribution (scores between user and item) over these paths and computed JS divergence between them. <- For the rule distribution of test phase path, all 1.0 are given?

Reference (for understanding)

• https://github.com/orcax/LOGER