借记
推论
计算机科学
因果推理
图形
水准点(测量)
人工智能
机器学习
理论计算机科学
数据挖掘
计量经济学
数学
心理学
大地测量学
地理
认知科学
作者
Rong Lin,Yongbin Liu,Cheng Ouyang
标识
DOI:10.1007/978-3-031-47240-4_18
摘要
The task of Knowledge Graph Completion (KGC) entails inferring missing relations and facts in a partially specified graph to discover new knowledge. However, the discrepancy in the targets between the training and inference phases might lead to in-depth bias and in-breadth bias during inference, potentially resulting in incorrect outcomes. In this work, we conduct a comprehensive analysis of these biases to determine their extent of impact. To mitigate these biases, we propose a novel debiasing framework called Causal Inference-based Debiasing Framework for KGC (CIDF) by formulating a causal graph and utilizing it for causal analysis of KGC tasks. The framework incorporates In-Depth Bias Mitigation to diminish the bias on feature representations by measuring the bias during inference, and In-Breadth Bias Mitigation to increase the distinguishability between feature representations by introducing a novel loss function. We evaluate the effectiveness of our proposed method on four benchmark datasets - WN18RR, FB15k-237, Wikidata5M-Trans, and Wikidata5M-Ind, achieving improvements of 2.5%, 0.9%, 3.2%, and 1.5% on Hit@1 respectively. Our results demonstrate that CIDF leads to significant improvements on these datasets, with more substantial gains observed in the biased settings on WN18RR achieving a 3.4% improvement in Hit@1.
科研通智能强力驱动
Strongly Powered by AbleSci AI