DeLTA seminar by Lei You

Speaker

Lei You from DTU

Title

Harnessing Transport Theory for PostHoc Explanations of Machine Learning

Abstract

In this talk, we explore advanced techniques in Explainable AI (XAI) by integrating concepts from optimal transport theory, a mathematical framework for comparing and aligning distributions. Two themes arecovered:

Finding Distributional Pattern as Explanations
Traditional counterfactual explanations focus on changing individual inputs to see how they affect outcomes, but they often miss the bigger picture of how groups of data points relate to one another. We extend traditional counterfactual explanations by introducing Distributional Counterfactual Explanation (DCE), which shifts from focusing solely on individual input changes to considering broader patterns within the entire data distribution. As a result, our approach provides stakeholders with valid counterfactual distributions supported by statistical confidence. 

Sparsifying Explanations With Minimal Actions

 We refine counterfactual explanations to enhance actionable efficiency by minimizing unnecessary feature changes, ensuring the proposed interventions are both valid and practical. Using optimal transport, we derive a joint distribution between observed and counterfactual data, which informs Shapley values for more precise feature attributions. This approach ensures minimal, realistic changes that make explanations more feasible and impactful for stakeholders.

Bio

Lei You, PhD, is an Assistant Professor in Applied Mathematics at the Technical University of Denmark (DTU). He received his Ph.D. in Computer Science, specializing in Mathematical Optimization, from the Department of Information Technology at Uppsala University in 2019. During his Ph.D., he interned as a visiting data scientist at The Boston Consulting Group (BCG) Gamma. Following his Ph.D., he worked as a data scientist at Bolt and Wolt (Doordash) in the domain of on-demand logistics optimization. His current research interests are centered around theories of interpretability and efficiency of machine learning models toward On-Device AI. He explores strategies to streamline complex models without performance loss and unravel the intricate mechanisms of decision-making models. Central to this pursuit is understanding the synergy between model simplification and explainability: reducing a model's complexity aids in elucidating its functions, and concurrently, explainability drives the efficient compression of the learning model.

Join the DeLTA community

You can subscribe to the DeLTA Seminar mailing list by sending an empty email to delta-seminar-join@list.ku.dk<mailto:delta-seminar-join@list.ku.dk>
DeLTA online calendar
DeLTA Lab page