Provably Better Explanations with Optimized Aggregation of Feature Attributions