Why educational data? - Education is inherently important and interesting, but it also has advantages from an ML perspective: available data is growing exponentially; many standard algorithms are either ineffective or uninterpretable on educational data; and finally, what space is more complex or has a higher intrinisic dimension than the human mind?

Why metacognition? - Metacognitive processes impact every educational domain, and thus pose an interesting ML problem: how do we learn models that will generalize to entirely new domains?

Why interpretation? - A model that predicts learning is good; a model that also generalizes to new data is better; a model with principles a human can understand is the best of all.

Why causality? - Predicting student performance is important, but unless we can also improve that performance, we're not really achieving the objective. Unfortunately, predictively optimal models may not be causally optimal.

Other... - I'm keenly interested, though secondarily, in behavioral predictors of affect. For example, can we build measures of motivation that don't ever explicity ask about motivation? In the long term, I'd also like to look at augmented learning, the use of every day experiences and activities as learning opportunities.