A Knowledge-guided Framework to Enhance Causal Reasoning and Human-AI Collaboration
Uzma Hasan, Md Osman Gani · 2025
Traditional Artificial Intelligence (AI) models often rely on correlations, leading to inaccurate conclusions and less effective recommendations. In contrast, Causal AI leverages data to uncover cause-and-effect relationships represented using a causal graph, enabling more reliable decision-making in complex systems. But observational data alone is often insufficient to uncover the underlying causal relations accurately, making the search for causal graphs computationally challenging due to an exponentially growing search space. Prior causal knowledge, such as expert-verified causal edges, can effectively guide the discovery process by restricting the search space to more accurate possibilities. Many fields, like healthcare, offer abundant prior knowledge through expert feedback, published research, and historical data. In this study, we introduce a Knowledge-Guided Causal AI System (KGS), which is a practical framework that incorporates domain-specific priors as structural constraints alongside observational data to improve causal graph discovery. KGS reduces the computational burden and enhances reliability by leveraging prior knowledge, optimizing the search process, and ensuring discovered edges align with established causal knowledge. Extensive evaluations on synthetic and benchmark datasets as well as on a real-world healthcare problem related to oxygen therapy treatment for ICU patients demonstrate its ability to improve discovery accuracy and efficiency. By reducing search complexity and enhancing trustworthiness, KGS can advance robust AI-driven decisions across critical domains like healthcare, autonomous systems, education, and finance. The resulting causal graphs can provide reliable and precise insights essential for decision-making, policy development, and enhancing human-AI collaboration across various domains.