What If We Had Played Another Song? Counterfactual Questions in Machine Learning, from Prediction and Intervention to Privacy

In this talk, Dimitri Staufer follows them through four questions, with a running example: wearables and personalised music for people living with dementia, where the body responds to every song in a way that a thumbs up never does:

1. Can we see it coming? Dimitri discusses interpretable time-series features built from local summaries, and what it takes to evaluate early warning rather than detection when the labels come from care records that may be late or incomplete.
2. What should we do now? Dimitri discusses why liking a song is not the same as being calmed by it, and, from our broader work on recommendation, how a learned kernel lets a compact model express two opposite tastes at once.
3. What if we had acted differently? Ongoing research on counterfactual forecasting asks what must stay the same between the history that happened and the one we imagine.
4. What if a person’s data had been left out? This is the same kind of question, asked about data, and differential privacy is a promise about its answer. Our work on language models separates recovering text from memorising it, and shows how a privacy guarantee bounds the training effect that any extraction or membership test can detect.

Across these questions, the comparisons we choose decide what our evidence can support: for the action we did not take, for the training data we did not use, and for the people whose data makes these systems work.