Research questionHow can we explain interacting reasons behind classifier predictions fast enough for repeated local use?Feature-attribution scores can obscure combinations of features that jointly influence a prediction. Generating detailed local explanations repeatedly can also be too costly for latency-sensitive applications.