Predoc seminar: Zhendong Wang

Seminar

Date: Tuesday 18 June 2024

Time: 13.00 – 15.00

Location: L30, DSV, Borgarfjordsgatan 12, Kista

Welcome to a predoc seminar on interpretability and trust when machine learning is used in healthcare. Zhendong Wang, PhD student at DSV, is the respondent.

On June 18, 2024, PhD student Zhendong Wang will present his ongoing work on “Constrained Counterfactual Explanations for Temporal Data”. The seminar takes place at the Department of Computer and Systems Sciences (DSV), Stockholm University.

Respondent: Zhendong Wang, DSV
Opponent: Pedro Pereira Rodrigues, University of Porto, Portugal
Main supervisor: Panagiotis Papapetrou, DSV
Supervisor: Isak Samsten, DSV
Professor closest to the subject: Rahim Rahmani, DSV

Contact information for Zhendong Wang

Find your way to DSV

 

Abstract

In machine learning, achieving high performance in predictive models while maintaining interpretability is crucial in critical domains, such as healthcare. Recent advances in algorithms in handling temporal data have gained success in time series prediction and event sequence modelling.

However, these algorithms remain opaque, which causes difficulty in understanding the prediction outcome and further leads to users not trusting the model. Counterfactual explanations, as a post-hoc explanation method, can provide actionable insights into machine learning models already deployed in different domains.

Specifically, counterfactual explanations can provide intuitive, sample-based explanations by suggesting modifications of the input sample to achieve the desired prediction outcome. There has not been much research on providing counterfactual explanations for temporal data predictions; hence, the thesis focuses on the counterfactual generation algorithms for time series classification, time series forecasting and event sequence classification. Furthermore, applying counterfactual generation requires domain knowledge to guide the counterfactuals to become more relevant.

Previous work has investigated plausible counterfactual changes but has also followed causal graph models to produce the counterfactuals. This thesis investigates applying two constraint mechanisms, local temporal and domain-specific constraints, to provide more relevant and efficient counterfactual explanations.