HERMES: Heterogeneous Effects Representation with Matched Embeddings using Siamese Networks

We consider the problem of estimating heterogeneous treatment effects from observational data. Specifically, we are interested in the estimation of conditional average treatment effects (CATE) functions, i.e. functions mapping the effect of a binary treatment to the space of unit-level covariates. In the absence of a controlled randomized mechanism of treatment assignment, simple comparisons between treated and control populations can be potentially confounded by significant distributional differences in the covariate space. In this context, recent representation learning strategies aim to learn balanced latent representations in a new space where the treated and control distributions are more comparable, reducing variance. We introduce HERMES (Heterogeneous Effects Representation with Matched Embeddings using Siamese Networks), a novel framework that integrates self-supervised contrastive learning into causal representation learning. HERMES employs a Siamese architecture that dynamically pairs individuals based on similarity in estimated individual treatment effects (ITE), encouraging representations where proximity reflects treatment-response similarity rather than covariate similarity alone. Unlike representation learning approaches that rely only on covariates, HERMES injects the ITE into representation learning, improving accuracy under standard assumptions. Experiments on IHDP and JOBS benchmarks show that HERMES improves the expected Precision in MSE by 14-15% over baselines, without added inference cost.