Dynamic Criminal Network Link Forecasting via Deep Reinforcement Learning

Deep Learning

Criminal network analysis (CNA) presents unique issues due to the clandestine nature of such networks, resulting in incomplete datasets with missing nodes (actors) and linkages (relationships). Traditional link prediction methods, often based on social network analysis (SNA) and supervised machine learning (ML) approaches, struggle to attain high accuracy in this domain due to their reliance on big datasets and static snapshots of network data. To address these limitations, we use deep reinforcement learning (DRL) to predict links in criminal networks. Our time-based DRL model (TDRL) learns from an evolving dataset. The TDRL model, which generates synthetic data through self-play or self-simulation, requires less data and adapts to dynamic network changes, outperforming traditional supervised methods like gradient boosting. Regarding forecasting accuracy. Our model uses natural language processing-inspired embedding approaches to improve computing efficiency during training..

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.