Telecommunication Fraud Detection via Improved Graph Convolution and Bidirectional Temporal Learning With Adaptive Fusion Strategy

ABSTRACT Telecommunication fraud has escalated in complexity due to evolving adversarial strategies that exploit dynamic communication patterns, multimodal signals, and semantic manipulation. Recent developments in deep learning and graph‐based modeling have shown promise; however, existing systems struggle to simultaneously capture temporal dependencies, relational feature structures, and linguistic nuances embedded in modern fraud activities. Addressing these limitations, this study proposes an Improved Graph Convolutional Network–Bidirectional LSTM (IGCN–Bi‐LSTM) framework integrated within a unified signal‐to‐text and multi‐perspective feature‐engineering pipeline for high‐accuracy fraud detection. The system begins by converting raw telecommunication signals into structured textual representations through a CNN‐driven signal‐to‐text encoder, enabling the extraction of temporal–spectral patterns. These sequences are subsequently enriched through a comprehensive feature‐engineering module that synthesizes linguistic markers, statistical descriptors, lexical indicators, and semantic embeddings. The hybrid IGCN–Bi‐LSTM model then jointly learns higher‐order relational dependencies among features and bidirectional temporal patterns, while an adaptive score‐level fusion mechanism optimally weights model outputs for robust classification. Experiments were conducted using a high‐quality synthetic Fraud Detection Transactions Dataset comprising 50,000 transactions with 21 heterogeneous attributes covering behavioral, contextual, financial, and security‐related characteristics. Extensive preprocessing, normalization, and stratified data partitioning ensured reliable training of the hybrid model in a GPU‐accelerated environment. The proposed model demonstrated substantial improvements over baseline methods by effectively capturing weakly correlated, high‐dimensional features and rare‐event patterns. Performance evaluation using precision‐recall metrics confirmed the superiority of the IGCN–Bi‐LSTM fusion, particularly in highly imbalanced scenarios where conventional accuracy metrics fail to reflect true detection capability.