Artificial Intelligence for causality with applications in the social and economic sciences
- Location: the accredited SeNSS DTP
- Duration: three years, beginning in October 2020 and completing in 2023
- Supervisor: Professor Paul Clarke
20 January: date to apply to ISER
17 February: date to apply for studentship, please submit this application to the supervisor(s) directly
16 March: final decisions communicated
We seek a talented student with a background in computer science, econometrics or statistics to work in one of the following areas or propose a relevant topic of their choice:
(i) Causal discovery for high-dimensional feature spaces: Most classic causal discovery methods take the covariate/treatment to be a high-level feature that can be manipulated by direct intervention or policy change. But some data sets only contain information on such low-level features - forming high-dimensional feature spaces - that the idea of manipulating these features is meaningless (e.g. a pixel in a video stream). The aim of this project is to use the latest techniques to identify manipulable high-level features from rich data on low-level features in order to carry out causal analysis using Artificial Intelligence techniques (e.g. reinforcement learning, causal inference) (Bengio et al. 2019).
(ii) Embedding common knowledge in causal effect identification: To understand the effect of a treatment without resorting to direct experiments requires us to incorporate knowledge about the social process; this should in turn allow us to learn a model for the interventional distribution from observational data. In this project, a body of common knowledge related to a topic from, e.g., sociology or economics, will be used to guide the creation of an interventional model using a variety of approaches, e.g. simulating scenarios to integrate with observed data or Bayesian methods.
(iii) Speeding up experimental designs by incorporating observational data: If available, abundant observational data can be used to help speed up the experimental process (e.g. Zhang et.al. 2017, Dudik et.al. 2014). The purpose of this project is to use advances in machine learning, causal inference and logic to identify from observational the further experimental interventions or data collection we need to carry out in order to make inferences about the causal effects of treatments.
2 Blog post on causal inference: https://www.inference.vc/untitled/
Common knowledge example: Speer, Robert, and Catherine Havasi. “Representing General Relational Knowledge in ConceptNet 5.” LREC. 2012.
General intro to causality: Pearl, Judea, and Dana Mackenzie. The book of why: the new science of cause and effect. Basic Books, 2018.
3 Zhang, Junzhe, and Elias Bareinboim. “Transfer learning in multi-armed bandits: a causal approach.” Proceedings of the 26th International Joint Conference on Artificial Intelligence. AAAI Press, 2017.
Dudík, Miroslav, et al. “Doubly robust policy evaluation and optimization.” Statistical Science 29.4 (2014): 485-511.