Structure-Preserving Counterfactual Regression for Personalized Consumer-Electronics Service Networks

Smart wearable devices increasingly form interconnected causal service networks to deliver personalized service-based health interventions, aiming to improve individual health outcomes. Most studies on Individual Treatment Effect (ITE) estimation in this context tackle treatment selection bias by minimizing the distributional discrepancy between control and treated groups using representation learning frameworks. However, structural relationships among users in smart wearable service networks are often inadequately preserved during distribution alignment and model learning, potentially reducing the accuracy of personalized causal service interventions. In this paper, we propose IC-CFR, a novel Counterfactual Regression approach explicitly designed for smart wearable causal service scenarios, preserving intra-graph and cross-graph structural information. IC-CFR introduces an intra-graph structure regularizer (IGR) for structurally consistent causal distribution alignment and a cross-graph structure regularizer (CGR) that maintains consistency between representation and covariate spaces. Extensive experiments on three widely used benchmark datasets demonstrate IC-CFR’s superior effectiveness compared to existing state-of-the-art methods in causal inference for service networks.