Predicting Crime Through Graphs: A New Paradigm in Drug Crime Analysis
In recent years, drug abuse and addiction have become a significant burden on society. Although substantial resources have been invested to meet public expectations for drug crime prevention, the rapidly changing social landscape has led drug offenders to seek new ways to evade investigations by producing, transporting, and distributing drugs through alternative regions. This has posed considerable challenges to effective drug prevention efforts.
To address this societal issue, we present a case study that utilizes drug crime networks and social network analysis(SNA) techniques to predict drug-related recidivism[1].
Drug-related recidivism can be predicted by leveraging network centrality measures
Background
Structural Limitations of Traditional Crime Prediction Models
Drug-related crimes often exhibit a highly organized nature and are frequently carried out through covert networks among offenders.
However, most crime prediction models to date have focused primarily on the demographic characteristics of offenders. As a result, models built using conventional statistical methods or based on individual attributes of criminals fail to capture the structural characteristics of criminal organizations, thereby showing limitations in accurately predicting recidivism.
Moreover, offenders often employ sophisticated evasion strategies, such as deliberately severing network ties to avoid detection. These behaviors are difficult to identify using models based solely on individual attributes.
Core Idea
Recidivism Prediction Using Crime Network Analysis
In this study, the strength of connections and centrality among drug-related offenders is quantified based on a drug crime network dataset.
These network-based indicators are then combined with traditional individual attributes of offenders and used as relational features.
Finally, a classification model is developed by training on both relational and individual attributes to predict the likelihood of recidivism.
Solution
Step 1. Construction of the Offender Network
This study collected drug-related crime data from the New Taipei City Police Department in Taiwan for the year 2017.
Offender Data (Nodes): Data was collected on 4,561 drug-related offenders, including six individual attributes such as gender, age, height, occupation, nationality, and education level.
Crime Network (Links): A network was constructed based on accomplice relationships among the 4,561 drug-related offenders
Step 2. Calculating Network Importance (Centrality)
To assess the importance of individuals within the relationships among drug-related offenders, four types of centrality analyses were conducted.
The calculated network centrality measures are added as individual attributes for each offender.
<Example Data>
※ Age and height are normalized to values between 0 and 1.
Step 3. Model for Identifying Potential Repeat Offenders
Using the training data described earlier, several offender identification models were constructed as follows.
Result
Recidivism Classification Model Reflecting the Structure of the Offender’s Accomplice Network
The machine learning model trained on both the basic information of drug-related offenders and their network centrality measures outperformed models trained solely on either type of data, achieving approximately 5 to 7 percent higher performance in recidivism classification.
Notably, the Decision Tree model showed the greatest improvement, with its performance increasing by 7.3 percent.
In conclusion, incorporating accomplice network structures in addition to basic offender information enables more accurate classification of recidivism.
<Performance Comparison by Training Data Type in Traditional Machine Learning>
※ F1-Score Results (%) for Recidivism Classification
Drug-related crime is no longer an individual issue but a product of complex relational networks. This study constructed a network based on accomplice relationships among offenders and analyzed each offender's centrality to predict the likelihood of recidivism. Graph-based features outperformed traditional individual attribute-based approaches, and the combined model using both achieved the highest accuracy.
Graph analysis is increasingly establishing itself as an effective tool for understanding and predicting complex crime patterns, and it is expected to become a core technology in the fields of crime prevention and security.
CYRAM Inc.
1660-4230 | netminer@cyram.com
References
[1] Fu-Ching Tsaia, Ming-Chun Hsub,c, Chien-Ta Chenc, Da-Yu Kaob (2019).
Exploring drug-related crimes with social network analysis. Procedia Computer Science, 159, (2019) 1907–1917.