HyKG-CF

A Hybrid Approach for Counterfactual Prediction using Domain Knowledge

authored by

Hao Huang, Maria Esther Vidal

Abstract

Predictive models are gaining attention as powerful tools for aiding clinicians in diagnosis, prognosis, and treatment recommendations. However, their reliance on associative patterns may raise concerns about reliability of decision support, as association does not necessarily imply causation. To address this limit, we propose HyKG-CF, a hybrid approach to counterfactual prediction that leverages data and domain knowledge encoded in knowledge graph (KG). HyKG-CF integrates symbolic reasoning (on knowledge) with numerical learning (on data) using large language models (LLMs) and statistical models to learn causal Bayesian networks (CBNs) for accurate counterfactual prediction. Using data and knowledge, HyKG-CF improves the accuracy of causal discovery and counterfactual prediction. We evaluate HyKG-CF on a non-small cell lung cancer (NSCLC) KG, demonstrating that it outperforms other baselines. The results highlight the promise of combining domain knowledge with causal models to improve counterfactual prediction.

Organisation(s)

L3S Research Centre

External Organisation(s)

German National Library of Science and Technology (TIB)

Type

Conference contribution

Pages

1104-1105

No. of pages

2

Publication date

10.03.2025

Publication status

Published

Peer reviewed

Yes

ASJC Scopus subject areas

Computer Networks and Communications, Computer Science Applications, Software

Sustainable Development Goals

SDG 3 - Good Health and Well-being

Electronic version(s)

https://doi.org/10.1145/3701551.3708813 (Access: Closed)