We present a machine-learning-based technique for selectively applying unsoundness in static analysis. Existing bug-finding static analyzers are unsound in order to be precise and scalable in practice. However, they are uniformly unsound and hence at the risk of missing a large amount of real bugs. By being sound, we can improve the detectability of the analyzer but it often suffers from a large number of false alarms. Our approach aims to strike a balance between these two approaches by selectively allowing unsoundness only when it is likely to reduce false alarms, while retaining true alarms. We use an anomaly-detection technique to learn such harmless unsoundness. We implemented our technique in two static analyzers for full C. One is for a taint analysis for detecting format-string vulnerabilities, and the other is for an interval analysis for buffer-overflow detection. The experimental results show that our approach significantly improves the recall of the original unsound analysis without sacrificing the precision.