Uncovering Factor Level Preferences to Improve Human-Model Alignment