🔹 Speakers: @jessicahullman.bsky.social, @doloresromerom.bsky.social, @tpimentel.bsky.social & Bernt Schiele

🕒 Call for contributions open until Oct 15 (AoE)
🔗 More info: eurips.cc/ellis

In short: Many XAI papers are based on goals such as "transparency". But what does that mean? We argue that XAI methods should be motivated by concrete goals (e.g., explaining how to change an unfavorable prediction) instead of vague concepts (e.g., interpretability).

Section 3, Misconception 1

Our article is also on arXiv: arxiv.org/pdf/2306.04292

Not that I know of. But the method is relatively easy to implement. Please reach out if you would like to use it. I'm happy to assist!

Sounds interesting? Have a look at our paper!

Joint work with Eric Günther and @ulrikeluxburg.bsky.social.

DIP is
✅ unique under mild assumptions,
✅ easy to interpret,
✅ entails an efficient estimation procedure,
✅ describes properties of the data (instead of just a specific model), and
✅ comes with a python implementation (github.com/gcskoenig/dipd).

In our recent AISTATS paper, we propose DIP, a novel mathematical decomposition of feature attribution scores that cleanly separates individual feature contributions and the contributions of interactions and dependencies.

Dependencies are not only a neglected cooperative force but also complicate the definition and quantification of feature interactions. In particular, the contributions of interactions and dependencies may cancel each other out and must be disentangled to be fully revealed.

For example, suppose we predict kidney function (Y) from creatinine (C) and muscle mass (M), and that C reflects Y but also M, which is not linked to Y. Here, M becomes useful once combined with C, as it allows us to subtract irrelevant variation from C. In other words, C&M cooperate via dependence!

Determining whether variables are relevant due to cooperation is crucial, as variables that cooperate must be considered jointly to understand their relevance. Notably, features cooperate not only through interactions but also through statistical dependencies, which existing methods neglect.