Analyzing User Patterns via Multi-Level Features for Telecommunications Fraud Detection

Telecommunications fraud poses a pervasive and serious threat, creating significant risks for both service providers and consumers. To mitigate these risks, we propose a Multi-Level Feature Extraction (MLFE) method that is different from conventional handcrafted feature engineering techniques. Our approach is designed to more effectively and correctly distinguish legitimate users from fraudsters by automatically extracting intra-user features across multiple call records with a set function. Additionally, we employ graph neural networks to enhance the characterization of inter-user relationships.We evaluate our method using real-world telecommunications data provided by a telecom network operator, demonstrating its high efficacy in accurately identifying high-risk users. Our findings reveal that high-risk users exhibit distinct patterns in terms of time, stability, region, and call targets. This research also provides valuable insights for the telecommunications company, which they would integrate into their fraud detection systems.