Counterfactual Causal Adversarial Networks for Domain Adaptation

Abstract

To eliminate domain shift in domain adaptation, most methods do so by encouraging the model to learn common features. However, the interpretability of these domain adaptation methods lacks in-depth research, and we note that the domain adaptation process can be regarded as a causal intervention, which can further form theoretical explanations in the causal relationship. Our proposed counterfactual causal adversarial network CCAN performs better on the domain adaptation task. Supported by causal theory, CCAN completes the adversarial learning of the network through counterfactual intervention, and uses the first proposed domain-adaptive causal effect to supervise the entire network. CCAN successfully validates the goal of evaluating the quality of domain adaptation through counterfactual intervention effects in causality to supervise the better completion of the entire task. The causal theory endows the whole CCAN with good interpretability. Experimental results on two challenging UDA benchmarks validate the superiority and effectiveness of CCAN for domain adaptation with counterfactual interventions based on causal theory, and analyze the role that domain adaptation causal effects play in the overall supervision.