The transaction records were chrono

The transaction records were chronologically split into trainingrecords and testing records. For parameter estimation and tuning,we only use labels of accounts appearing in the training records,while performance was evaluated on accounts appearing in thetesting records. In testing we paid special attention to the "missedbad": the Fraud-labeled accounts that were not evaluated as riskyby the Risk Detection System.In our experiment, the Community Detection module was pro-vided with a list of 90401 Risky-labeled accounts as seeds, returning71074localcommunities.Notethatnotallseedssuccessfullyyieldedcommunity because of the conductance requirement.