6.8 Chicago Internal Affairs Uses Neural Network: A Case Study

The Chicago Police Department's (C.P.D.) Internal Affairs Division used neural networks to study 200 officers who had been terminated for disciplinary reasons and developed a model for predicting current officers with a likelihood of having similar disciplinary problems. The model, when compared to current department officers, produced a list of officers who it determined were "at risk" of having some future problems

The C.P.D. Internal Affairs Division used the model to study the records of 12,500 current officers. These personnel records included such information as age, education, sex, race, number of traffic accidents, reports of lost weapons or badges, marital status, performance reports, and frequency of sick leaves. The model was able to produce a list of 91 at-risk officers. Of those 91 people, nearly half were found already to be enrolled in a counseling program founded by the personnel department to help officers guilty of misconduct. The Internal Affairs Division wanted to use the neural network model to supplement the counseling program, because the sheer size of the Chicago police force makes it nearly impossible for all at-risk officers to be identified by their supervisors.

The motivations cited by the developers for wanting to use neural networks was that the software could be effective for two reasons: (1) they observed that as the number of variables increased, so did the accuracy of the predictions, and (2) they found that neural networks could effectively deal with missing data, which for this application was often the case, as some of the personnel files contained text narratives and were not uniform. This is an important feature of neural networks: their ability to perform with incomplete data. The importance of training the networks with good examples should also be noted.

Despite the ethical discussion raging over whether a neural network should be used to monitor human beings, an issue raised by the brotherhood union, the model can not be accused of being subjective and personally biased, as can human-based evaluations. Clearly, the software can hold no personal grudges and seeks only to identify patterns and examine behavioral characteristics that could spell trouble. The alternative system, being human-based, cannot avoid subjectivity and bias at some level. It is worth noting that the Fraternal Order of Police "vehemently opposed" the department's old system for that very reason.

To counterbalance the inherent "dispassion" of the neural network, the department closely examined the software's findings to ensure that officers who are clear anomalies, and thus don't warrant being on the list, are removed from consideration. This combination of objective technology and subjective humanity does not necessarily spell perfection, but it demonstrates that hybrid systems that incorporate machine and human intelligence are clearly the optimal methodology for investigative data mining; care, however, must be taken as, often, the use of data mining will raise the issues of privacy and human rights.