ColdNet: Treatment effect estimation with cold-start, imbalance, and zero-inflated outcomes

Individual treatment effect (ITE) estimation from observational data becomes unreliable when three challenges co-occur: extreme class imbalance (0.4% treatment rate), outcome sparsity (97.6% zeros), and pervasive cold-start (99.2% incomplete profiles). These conditions violate identifying assumptions—propensity scores collapse toward boundary values, and outcome predictions degrade for subjects with sparse historical features. We present ColdNet, a neural causal architecture with three innovations: (1) outcome-stratified ensemble learning that reduces effective imbalance from 1:256 to 1:2 while preserving outcome heterogeneity; (2) targeted regularization with sparsity-aware preprocessing that forces balanced representations via counterfactual correction; and (3) cluster-based cold-start enhancement that transfers predictions from similar training samples via locality-preserving quantile aggregation. On a production e-commerce dataset for 3P seller recommendations (1.53M training, 590K test samples), ColdNet achieves 27.6% MAE and WAPE improvement on cold-start cases and 82.8% median error reduction, while semi-synthetic validation shows 13.9× better treatment effect estimation than Double Machine Learning under identical imbalance. ColdNet is deployed in production, processing 4 Billion+ predictions weekly in US and 3 EU Marketplaces currently.