Data-Driven Bayesian Network for Risk Analysis of Telecom Fraud

Abstract

In light of the prevalent nature of global telecom fraud, proactive crime prevention and control are increasingly imperative. This paper introduces a novel hybrid framework for fraud risk analysis using Bayesian network, based on a data-driven approach and skillfully integrating D-S evidence theory to incorporate prior background knowledge. By analyzing data collected from existing cases, the study summarizes and identifies risk-influencing factors from multiple dimensions. We compare three structure learning algorithms: PC, Bayesian Search, and the Greedy Thick Thinning (GTT), to explore the underlying relationships between risk-influencing factor and successfully construct a robust Directed Acyclic Graph (DAG). The Expectation-Maximization (EM) algorithm is employed to accurately estimate network parameters and conditional probability distributions. The results demonstrate that our approach can automatically discover causal links between fraud risk factors and handle the nonlinear complexities of multiple factors. Finally, through scenario reasoning, sensitivity analysis, and validation of model correctness, the effectiveness and rationality of the Bayesian network are further affirmed. This research provides public security organs with a potent data support and decision-making tool to prevent and control telecom fraud crimes.