ساخت شبکه های بیزی برای پرونده سازی کیفری از داده های محدود

The increased availability of information technologies has enabled law enforcement agencies to compile databases with detailed information about major felonies. Machine learning techniques can utilize these databases to produce decision-aid tools to support police investigations. This paper presents a methodology for obtaining a Bayesian network (BN) model of offender behavior from a database of cleared homicides. The BN can infer the characteristics of an unknown offender from the crime scene evidence, and help narrow the list of suspects in an unsolved homicide. Our research shows that 80% of offender characteristics are predicted correctly on average in new single-victim homicides, and when confidence levels are taken into account this accuracy increases to 95.6%.

The study of criminal behavior for the purpose of identifying the characteristics of an unknown offender and the motivation for the crime is commonly known as criminal profiling. In current practice, criminal profiling relies primarily on the personal experience of criminal investigators and forensic psychologists, rather than on empirical scientific methods [31]. As such, it may be subject to errors caused by cultural biases and misinterpretation [24], [31], [32] and [43]. After clearing a criminal case, investigators file the background characteristics and psychological diagnosis of the convicted offender together with the forensic evidence obtained from the crime scene. With the increased availability of computer and information technologies, law enforcement agencies have been able to compile databases with detailed offender and crime scene information from major felonies, such as murder, rape, and arson. Consequently, important authors have advocated that machine learning techniques will play a significant role in developing decision-aid tools for police investigations [4], [17], [27], [32] and [42]. The most significant contributions to date have been recently reviewed in [17]. Rule-based systems have been proposed in [4] for knowledge acquisition from a database with modus operandi information. Research on inductive profiling has employed statistical analysis to classify offender behavior into categories or dichotomies, based on the crime scene evidence [12], [25], [34], [35], [37], [39] and [41]. While this research has been successfully implemented to predict the approximate residence location of serial homicide offenders [35], it has been unable to identify psycho-behavioral offender profiles in single-victim (non-serial) homicides. This shortcoming has been attributed to the complexity of human behavior and to the large number of relevant variables, both of which limit the applicability of behavior classification techniques [2], [31] and [32]. In this paper, a novel Bayesian network (BN) approach to criminal profiling is presented. The approach consists of learning a BN model of offender behavior from data and, subsequently, implementing the model for profiling by means of an inference engine. The database used in this paper is similar to the modus operandi database described in [4]. However, the BN approach is not limited by decisive “if-then” relationships, because it views the relationships among all variables as probabilistic. Unlike inductive profiling, the BN approach does not require to postulate behavior categories a priori and, consequently, it is capable of identifying psycho-behavioral profiles in single-victim single-offender homicides (Section 6). Also, the inferred offender characteristics include confidence levels that represent their expected accuracy. Thus, when provided with a BN profile, the police can easily establish what are the reliable predictions in the investigated case. Implementing BN models for inference has proven valuable in many applications, including medical diagnosis, economic forecasting, biological networks, and football predictions [1], [19], [20], [23] and [30]. This literature shows that the effectiveness of BN inference and prediction is highly dependent on the sufficiency of the training database. While various approaches have been proposed for dealing with insufficient databases [11], [14], [15], [16], [21], [26], [30] and [40], there are no general guidelines for establishing whether a given database is insufficient. In [45], it was shown that the size of a sufficient database depends on the number of variables, their domain, and the underlying probability distributions. But, while the variables and the domain definitions are known from the problem formulation, the underlying probability distribution is often unknown a priori. This paper presents a set of performance metrics that can be used to determine the sufficiency of an available database without knowledge of the underlying joint probability distributions (Section 4). Although a police database may include hundreds of cleared cases, they may still be insufficient to train a BN model due to the large number of relevant variables, and to the complexity of their relationships [3]. Therefore, in Section 5 these performance metrics are implemented to determine the size of a sufficient database with single-victim single-offender homicides. Subsequently, a BN model is trained using a newly modified K2′ algorithm that improves performance once the database size is fixed (Section 5). In Section 6, the trained BN model is applied to infer the characteristics of unknown offenders from the crime scene evidence. The results show that when the confidence level is taken into account, the average accuracy of the BN predictions is 95.6%. For comparison, the evidence from two homicide cases has been presented to a team of expert criminologists. Based on the evidence alone, the experts predict 53% of all offender variables correctly. Whereas, in the same two cases the BN predicts 86% of all offender variables correctly, and displays 80% average accuracy in 1000 other homicide cases. Also, offender characteristics that cause disagreement among the experts are predicted correctly and with a high confidence level by the BN. Finally, the structure of the BN model indicates what are the most significant relationships among the variables and, thus, it could be used for the scientific development of hypothesis on criminal psychology.

The increased availability of computer and information technologies has enabled law enforcement agencies to compile extensive databases with detailed information about major felonies, such as murder, rape, and arson. Consequently, several authors have advocated that machine learning techniques will play a significant role in developing decision-aid tools for police investigations [4], [17], [27], [32] and [42]. The most significant contributions to date have been recently reviewed in [17]. In this paper, we develop an approach for obtaining BN decision-aid tools that consists of the following steps: (1) assessing the sufficiency of an available database; (2) training a BN model using both expert knowledge and data; and, (3) implementing an inference engine to produce offender profiles in unsolved cases. Numerical studies demonstrate that the BN model can be used to successfully infer the characteristics of an unknown offender from the crime scene evidence in single-victim homicides. On average, 80% of the offender characteristics are predicted correctly by the BN profile. Moreover, since each prediction is accompanied by a confidence level that is proportional to its expected accuracy, by considering only predictions with high confidence levels the average accuracy increases to 95.6%. Hence, the BN profile can be implemented by investigators to narrow the list of suspects in unsolved homicides, and identify the motivation for the crime.