We present a novel approach to visualization recommendation that monitors user behavior for implicit signals of user intent to provide more effective recommendation. This is in contrast to previous approaches which are either insensitive to user intent or require explicit, user specified task information. Our approach, called Behavior-Driven Visualization Recommendation (BDVR), consists of two distinct phases: (1) pattern detection, and (2) visualization recommendation. In the first phase, user behavior is analyzed dynamically to find semantically meaningful interaction patterns using a library of pattern definitions developed through observations of real-world visual analytic activity. In the second phase, our BDVR algorithm uses the detected patterns to infer a user’s intended visual task. It then automatically suggests alternative visualizations that support the inferred visual task more directly than the user’s current visualization. We present the details of BDVR and describe its implementation within our lab’s prototype visual analysis system. We also present study results that demonstrate that our approach shortens task completion time and reduces error rates when compared to behavior-agnostic recommendation.