Answering Counterfactual Queries on Graph Databases

Hsi-Wen Chen, Jian Pei, De-Nian Yang, Ming-syan Chen

Keywords: Counterfactual Analysis, Graph Database

TL;DR: CF-GDB supports counterfactual reasoning on real graph database through C2GQ, leveraging CFI and CSI indices to deliver efficient and accurate search.

Abstract: Counterfactual analysis on graph data is central to causal reasoning and interpretability, yet existing graph-based methods rely on ad hoc perturbations and remain tied to model behavior rather than underlying data. To address this challenge, we introduce Counterfactual Graph Database (CF-GDB) queries, the first query-based framework for counterfactual reasoning on graphs that grounds counterfactuals in verifiable database instances. Our approach abstracts graphs into semantically meaningful concepts and compares them using a hypergraph-based distance that integrates local structure with global semantics. To ensure efficiency and scalability, we propose two complementary indices: the Concept Distribution Index (CDI), a histogram that provides certified lower bounds, and the Concept Semantic Index (CSI), a continuous embedding that provides upper bounds. These indices yield provably tight sandwich guarantees and enable efficient candidate pruning while preserving the fidelity of counterfactual retrieval. Using 8 read data sets across 4 domains, CF-GDB improves accuracy by over 20% and achieves up to 20× faster performance, demonstrating both fidelity and scalability.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 10171