Detecting pickpocketing gangs on buses is critical for safety and public security department. Knowing this both in real time and from historical records would allow effective law enforcement and crime prevention. However, very little research has been devoted into identifying pickpocketing gangs in an automated and holistic manner. This research utilizes smart card data generated by bus riders to identify pickpocketing gangs, who possess distinct characteristics from regular passengers. Particularly, we create a dataset of 1,098 pickpockets among 4.06 million bus riders in August, 2015 in Beijing automatically and efficiently based on an efficient labeling model of outliers. This model examines anomaly of passengers using the so-called relative outlier cluster factor and local outlier factor. The proposed mobility patterns of pickpockets are then learned based on supervised classification. Pickpockets from the derived dataset form a pickpocketing network, which is modeled as a graph with vertices denoted as discrete pickpockets, and edge weight quantified by a combined similarity on mobility pattern, space and time. A graphbased Louvain algorithm is adopted to detect pickpocketing gangs. Experiments are conducted on SINA microblog data to verify the detected pickpocketing gangs identified by the proposed framework. Results show that the framework detects 63 pickpocketing gangs and verifies 34 gangs by microblogs, with recall value 0.85. Findings from this research can assist police and public safety departments in the city in taking pro-active actions to track down pickpocketing gangs.