At this seminar you will hear from Amit Sharma who  is a Principal Researcher at Microsoft Research India. His work bridges causal inference techniques with data mining and machine learning, with the goal of making machine learning models generalize better, be explainable and avoid hidden biases.

Abstract

As predictive models based on machine learning (ML) are applied in diverse domains from healthcare to finance, reliability of the black-box models is one of the biggest problems facing ML research today: would the model generalize to out-of-distribution data, would it be fair to different subpopulations, and would it be explainable?

Amit will show how incorporating causality in predictive models can be useful for all these problems. Standard causal inference methods like matching can be applied to make deep learning models generalize to unseen distributions. A simple method based on matching obtains state-of-the-art results on domain generalization benchmarks. Similarly, the concept of counterfactuals is essential to explain machine learning models. Amit will describe work on automatically finding counterfactuals for any prediction and using it to create counterfactual-based explanations.

Finally, Amit will show that causal graphs are fundamental to understand the fairness of a ML system. Common fairness metrics fail to ensure fairness unless they account for the causal process that generates a given dataset.

Economic Data Science Seminar Series

The Alan Turing Institute, the Data Science Campus of the Office for National Statistics (ONS), Centre for microdata methods and practice (Cemmap) and the Financial Conduct Authority (FCA) are collaborating for the economic data science seminar series. The intended audience for these seminars includes both policymakers and academics, and the seminars will be overseen by a group of seminar partners made up of representatives of key stakeholder organisations.

The economic data science series has four objectives:

• to promote economic data science as a discipline
• to link researchers across institutions and disciplines and bring together the academic and policy communities
• to demonstrate cutting-edge applications of data science to economic questions to an audience of economists, data scientists and economic data scientists
• to provide a forum for addressing challenges in current work.

This session will be recorded.

Please be aware that once your registration is confirmed, we will forward a link onto you to access the webinar 48 hours before the event.