Less is More: Improving LLM Alignment via Preference Data Selection