Telecom Fraud Detection via Dual Hypergraph Neural Network under Multi-Dimensional Sparsity

With the increasing sophistication and concealment of telecommunication fraud, traditional detection technologies face significant challenges. Graph Neural Networks (GNNs) offer a new paradigm for addressing these challenges by concurrently modeling user nodes and their communication links to learn highorder representations of latent interaction patterns. However, their performance in practical applications is severely constrained by the ”multi-dimensional sparsity” inherent in telecommunication data: structural sparsity (missing relations), sample sparsity (extreme class imbalance), and semantic sparsity (underutilized edge attributes). To address these issues, this paper proposes a telecom fraud detection model based on a Dual Hypergraph Neural Network. First, a relationship-aware graph structure enhancement mechanism is proposed. It reconstructs latent homophilic links via behavioral similarity to mitigate structural sparsity, and subsequently extracts high-quality discriminative subgraphs by evaluating relationship strength through communication intimacy and Term Frequency-Inverse Document Frequency (TFIDF) measures. Second, to tackle sample sparsity, an improved data augmentation algorithm incorporating a graph structure expansion mechanism is designed, utilizing feature space interpolation and decoder-based structure reconstruction to mitigate class imbalance. Finally, a Dual Hypergraph Transformation (DHT) mechanism is introduced to map communication edges to hypergraph nodes, enabling the explicit modeling of high-order communication semantics. Experimental results on the test set demonstrate that the proposed model successfully overcomes the zero recall dilemma that plagues traditional baseline models when labels are extremely scarce. By explicitly modeling higher-order interactions, the model achieves a macro F1 score of 46.90% and a fraud recall rate of 69.90%. Compared with other existing approaches, the proposed method shows strong capability in identifying fraudulent users under the challenging conditions of sparse and imbalanced data.