Abstract: It's one of the social issues, that the different authorities in the world pay considerable attention to the ex-convicts for recommitting crimes. Previous studies of ex-convict classification are usually focused on static attributes of historical criminal data rather than dynamic trajectory. And fewer are focused on risk predicting analysis of crime-recommitting of larceny ex-convicts. For this reason, a larceny ex-convict (first offenders and recidivisms) classification was studied by combining static attributes and dynamic trajectory in this paper. Firstly, a long-timespan data base about the larceny ex-convict classification was developed based on static attributes and dynamic trajectory. Then the performances of different types of machine learning models on larceny ex-convict classification were explored and compared to define the most relevant features to it. Finally, an early warning model of larceny crimes was established based on weighted association rules. The research achievement can be applied to early warning of larceny crimes, and will show practical significance for crime crackdown and security precaution.