Workshop format

The workshop was split into two days: an in-person workshop in Jeju on 5 August, and a virtual day on 15 August for those authors who could not make it to Jeju. There were 13 papers at the in-person day, 12 online. 

The in-person workshop was well attended, with over 100 people registered, and more than 80 people in each session throughout the day.

For those interested, Tim live-Tweeted the in-person workshop, with a short summary of every presentation

Keynote

At the in-person day, Professor Elisabeth André (left) gave an excellent talk titled Presenting Users with Alternative Realities: Counterfactual, Semi-Factual, and Alterfactual Explanations in XAI, which demonstrated the benefits of combining disciplines, in this case, AI, HCI, and cognitive science.

Poster session

Our poster session was also a success again – so much lively discussion that gave attendees a chance to talk in detail with authors about their work. As well as workshop authors, we invited authors of papers related to the workshop to come along and present.

Some high-level thoughts

Some high-level thoughts on the current states of the field based on observations from the workshop papers:

1. It is clear that the field has matured a lot in the last 5 years. At the first instance of the workshop in 2017, many of the presentations were on algorithms that could highlight feature importance or some other information, with minimal consideration of whether this would be understandable, there were very few (if any?) human studies, and little focus on what the goal of the algorithms were: understandability, assistance on a downstream task, etc. In the 2024 edition, there were studies using co-design approaches with stakeholders, much more interest in the goals of methods, and several applied studies.

2. Similar to the above, there is a much greater focus on evaluation (see right some survey results from the online day). Not just papers proposing evaluation, but many authors had clearly thought hard about their evaluation, what it means, and how it relates to their goals. New metrics were being discussed, with in-depth ideas about how they relate to other metrics.

3. There is much less focus on perturbation-based feature attribution than previously – this having been identified as a source of error and instability, so should be used carefully.

4. Counterfactuals (and other *-actuals, such as alter-factuals and semi-factuals) are still having success in human behavioural studies, providing further evidence that these techniques do have some cognitive merit.

Until next year!

Tim, Mor, Tobias, and Hendrik

(And big thanks to Marija Slavkovik for giving us the idea of 'publishing' a reflection with our proceedings).