Learning Transparent Reward Models via Unsupervised Feature Selection