In this talk, I will share my experience of research on data-driven static analysis. An ideal static analysis would adapt to a given analysis task automatically, avoiding the unnecessary application of techniques that only increase the analysis cost without enhancing the precision. However, developing a cost-effective adaptation strategy for real-world programs has been challenging; designing such a strategy has been done by trials and error, requiring a huge amount of manual effort and expertise. Furthermore, such hand-tuned strategies are suboptimal and brittle. Our approach to overcome this shortcoming is to combine static analysis and machine learning, where the adaptation strategies are automatically learned from codebases without reliance on analysis designers. Toward this goal, we have developed machine learning models, efficient learning algorithms, and automated feature-engineering techniques appropriate for the static analysis application. I will talk about the overall approach, current achievements, and remaining challenges.