Multimodal Trilemma GNN‐LLM‐XAI Framework Resolving Accuracy Transparency Comprehensibility in E‐Commerce Fraud Detection

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

E‐commerce platforms face an escalating challenge from sophisticated financial fraud and cyberattacks, yet current detection systems suffer from a fundamental trilemma: achieving high detection accuracy, providing expert‐level transparency, and ensuring user comprehensibility simultaneously. Graph neural networks (GNNs) excel at detecting complex fraud patterns through relational learning but operate as black boxes, while existing explainable AI (XAI) methods generate technical attributions that overwhelm non‐expert users. Standalone large language models (LLMs) can produce natural language explanations but lack access to relational graph structures and risk scoring mechanisms. We present MT‐GXL, a novel architectural framework that resolves this trilemma through triple‐granularity orchestration of multi‐modal GNN‐based detection, XAI attribution, and grounded LLM generation. Our framework introduces three key innovations: (1) a triple‐granularity XAI router that dynamically routes explanations based on threat taxonomy; (2) a fraud‐aware homophily index that quantifies closed‐loop fraud ring cohesion; and (3) a grounding mechanism with claim verification that prevents LLM hallucination. Extensive experiments across three benchmark datasets demonstrate that MT‐GXL achieves 97.8% accuracy (AUC = 0.962) on IEEE‐CIS and 98.9% accuracy (AUC = 0.991) on CIC‐IDS2017; on the extremely imbalanced PaySim benchmark (0.13% fraud), where raw accuracy is uninformative, it attains F1 = 0.950 and AUC‐PR = 0.976. Crucially, MT‐GXL is the first unified system to simultaneously satisfy all three trilemma objectives: high accuracy (Acc 0.95), expert transparency (Trans 0.85), and user comprehensibility (Comp 4.0/5.0). A human‐centric evaluation with 70 participants confirms that LLM‐driven explanations significantly outperform traditional XAI, achieving 47% improvement in clarity, 71% in actionability, and 119% in anxiety reduction (). Computational analysis confirms production readiness: 110 ms inference latency, linear scaling to 10M transactions, and a 95% reduction in LLM API cost through tiered deployment.