Data Distribution Matters: Enhancing model generalization in Data-Driven Materials Discovery by intentionally exploiting both positive and negative data