A Bayesian network (BN) model of criminal behavior is obtained linking the action of an offender on the scene of the crime to his or her psychological profile. Structural and parameter learning algorithms are employed to discover inherent relationships that are embedded in a database containing crime scene and offender characteristics from homicide cases solved by the British police from the 1970s to the early 1990s. A technique has been developed to reduce the search space of possible BN structures by modifying the greedy search K2 learning algorithm to include a-priori conditional independence relations among nodes. The new algorithm requires fewer training cases to build a satisfactory model that avoids zero-marginal-probability (ZMP) nodes. This can be of great benefit in applications where additional data may not be readily available, such as criminal profiling. Once the BN model is constructed, an inference algorithm is used to predict the offender profile from the behaviors observed on the crime scene. The overall model predictive accuracy of the model obtained by the modified K2 algorithm is found to be 79%, showing a 15% improvement with respect to a model obtained from the same data by the original K2 algorithm. This method quantifies the uncertainty associated with its predictions based on the evidence used for inference. In fact, the predictive accuracy is found to increase with the confidence level provided by the BN. Thus, the confidence level provides the user with a measure of reliability for each variable predicted in any given case. These results show that a BN model of criminal behavior could provide a valuable decision tool for reducing the number of suspects in a homicide case, based on the evidence at the crime scene.
展开▼
