A key issue in designing a focused Web crawler is how to determine whether an unvisited URL is relevant to the search topic. Effective relevance prediction can help avoid downloading and visiting many irrelevant pages. In this paper, we propose a new learning-based approach to improve relevance prediction in focused Web crawlers. For this study, we chose Naive Bayesian as the base prediction model, which however can be easily switched to a different prediction model. Experimental result shows that our approach is valid and more efficient than related approaches.