On Causal Discovery and Inference from Observational Data
- Publication Type:
- Thesis
- Issue Date:
- 2019
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Causality is a fundamental component in all fields of science. In contrast to associational dependencies that are widely used in existing predictive machine learning and data-mining methods, causality implies the mechanism of how variables take their values and how the change of causes would lead to the change in the outcome. In the era of big data, for scientific discovery and rational decision-making, we fundamentally need methods for learning causal relationships between variables and estimating causal effects from observational data.
In this thesis, we aim to develop new models and algorithms for learning causal relationships and estimating causal effects using observational data. In particular, for the purpose of modelling and learning causal relationships from observational data, we study dynamic causal systems with feedbacks. To overcome the weakness of existing models that are unable to model both instantaneous and cross-temporal causal relations simultaneously, we propose a First-order Causal Process (FoCP) model and a causal structure learning algorithm to learn the causal graph of FoCPs from time series. For the purpose of estimating treatment effects, we investigate a range of existing methods for causal effect estimation, and propose three new methods using advanced machine learning techniques. First, to relieve the high-variance issue of the classic Inverse Propensity Weighting (IPW) estimator and thus to get more stable treatment effect estimates, we reframe it to the importance sampling framework and propose a novel Pareto-smoothing method using the generalized Pareto distribution from the extreme value statistics. Second, for causal inference with unobserved confounders, we take advantage of proxy variables and use deep latent variable models to model the underlying data-generating process. Building on recent advances in Bayesian inference and deep generative models, we propose a Causal Effect Implicit Generative Model (CEIGM). Finally, with an observation that most of existing methods for causal inference are essentially indirect in that they estimate the target treatment effects by first estimating other auxiliary quantities, we propose the idea of direct treatment effect estimation. Based on this idea, we further propose two deep neural networks for direct treatment effect estimation.
We evaluate all the methods proposed in this thesis using simulated, semi-simulated or real-world data. Experiment results show that they perform generally better than their competitors. Given the key importance of learning causality and causal inference in both theory and real-world applications, we argue that our proposed models and algorithms are of both theoretical and practical significance.
Please use this identifier to cite or link to this item: