Subset to interesting training data

[lboschman]

Currently the model trains on the entire data set, which consists mostly of regular events. Maybe we can daisy-chain models to first filter for potentially interesting data points, and then train a second model on those?