MUNet: A Multi-outcome Uplift Network for Modeling Multi-dimensional Treatment Effects

Uplift modeling estimates individual-level causal effects, yet most existing methods address only a single outcome, overlooking the fact that real-world interventions often influence multiple correlated outcomes. Independent modeling fails to exploit shared causal structure and may produce inconsistent estimates. We propose MUNet, a unified multi-outcome framework for multi-dimensional uplift estimation. MUNet learns shared causal representations while enforcing additive consistency between category-level and total uplifts. The extended MUNet-X incorporates content-based gating and outcome-wise uncertainty weighting to model sample-dependent outcome relations and handle heterogeneous noise. Experiments on controlled synthetic data and the Dunnhumby Complete Journey dataset show that MUNet consistently outperforms representative single-outcome uplift methods, achieving higher ranking quality and stronger structural coherence. Repeated-split evaluation further indicates reduced variance for MUNet-X, demonstrating improved generalization under observational noise. These results highlight the value of structured multi-task modeling for realistic uplift applications.