PhD Seminar • Artificial Intelligence | Explainable AI • CREDENCE: Counterfactual and Rule-Based Explanations for Document Rankers

Wednesday, November 20, 2024 12:00 pm - 1:00 pm EST (GMT -05:00)

Please note: This PhD seminar will take place in DC 1304.

Joel Rorseth, PhD candidate
David R. Cheriton School of Computer Science

Supervisor: Professor Lukasz Golab

In pursuit of enhanced explainability for modern information retrieval systems, we present CREDENCE, an interactive tool designed to explain rankings generated by document ranking models. Specifically, CREDENCE explains the role of queries and documents in individual ranking predictions. We propose novel counterfactual explanation formulations that show how query and document perturbations lead to different rankings. Furthermore, we propose rule-based explanations that consolidate our counterfactuals and summarize larger trends. Our explanations are implemented in a user-facing tool that generates explanations in real time, and allows users to test their own query and document perturbations. Ultimately, CREDENCE facilitates recourse for undesirable ranking outcomes, while helping users understand the behavior of ranking models.