Analysis of a Few Domain Adaptation Methods in Causality

Understanding the effect of intervention is of great importance in many domains such as marketing, governance, health, economics, social science, etc. An ideal approach for estimating the effect of intervention requires conducting experiments which are often unethical, expensive, time consuming, or even impossible; leaving interesting business and research questions un-answered. Nowadays, data from businesses, government databases, and electronic medical records are generated in large amount and at unprecedented rate, making the use of observational study a viable alternative. However, the data posses bias between the treated and the control subjects posing a great challenge for this task. Machine Learning (ML) and Deep Learning (DL) models are recently deployed for causality and have achieved a state of the art results. "Correcting" the bias through aligning the distribution of the treated and control in the form of domain adaptation is shown to be an effective technique for estimating causal parameters. However, most often, these models involve complex DL architectures. There are tons of Domain Adaptation (DA) methods developed to align the shift between the source and target distribution in classical ML. In this thesis, following the Potential Outcome framework with binary treatment setting, we bring the idea of correlation alignment methods, adversarial training, and a parallel two streams architecture from domain adaptation into causality. But we initially built simple baseline Neural Networks (NN) models in each case which are optimized and evaluated. This is to understand the effectiveness and performance of the simple models in causality without any form of distribution alignment mechanisms proposed in domain adaptation literature. Then the DA components of these models are incorporated as an additional loss to the baseline models in each case, and are evaluated on four most widely used benchmarks. Our results show that incorporating additional DA losses are generally not effective for causality. The simple baseline models were able to achieve state-of-the-art results on some metrics. Suggesting that DL models with hyperparameter tuning could estimate causal parameters without necessarily the need for specialized regularizations. Moreover, many of the metrics could be estimated effectively with linear versions of these models. It was found that no method is superior over others on all the datasets. However, methods based on shared weights have fairly performed better than model based on unshared weights. Further more, using geodesic and Euclidean distances for correlation alignments produced similar results, implying some robustness to distance measure.