Deep Research‑Driven Risk Root‑Cause Analysis with Domain Graph Constraints for Large‑Scale Advertising Traffic

Abstract: In massive and complex advertising traffic, risk clues are sparse and hidden. Traditional methods struggle with efficiency, effectiveness, and controllability. The Alibaba‑Mama risk‑control team proposes an intelligent analysis mechanism that equips large models with a "domain‑graph" navigation system to guide structured exploration and rational judgment, aiming for explainable and responsible risk control.

Demo Overview: Using an "advertiser complaint handling" scenario, the core workflow demonstrates how the system receives a product ID and date range, automatically plans an analysis scheme, and iteratively judges whether traffic is fraudulent, ultimately generating a structured report and triggering compensation.

Background: Alibaba‑Mama serves millions of advertisers with billions of dollars in ad spend, making it a prime target for black‑gray market attacks. Existing risk‑control relies heavily on expert knowledge, which is limited by expertise scarcity, high cost, and instability, especially against sparse, context‑poor traffic.

Challenges and Opportunities: Recent advances in general LLMs (GPT, Claude, Qwen) offer knowledge generality, multimodal understanding, and prompt‑driven reasoning, opening possibilities for unsupervised exploration in risk control. However, key difficulties remain: uncontrolled planning leading to unstable results, and hallucinations causing logical errors.

Design Philosophy: The proposed DeepString RR (Reasonable & Responsible) system integrates three capabilities—Task Planning, Tool Selection/Calling, and Response Generation—mirroring the Deep Research paradigm. It combines dynamic domain‑graph constraints with LLM reasoning to achieve high effectiveness, efficiency, and determinism (precision > 90%, recall ≈ 99.9%).

Deep Research Mechanism in Risk Control: The system performs multi‑hop graph search to expand sparse contexts into dense sub‑graphs, then leverages LLM reflection and self‑critique to prune paths, preserving essential information while compressing tokens. Multimodal node descriptions (text, images, statistics, temporal behavior) are fed via system prompts to guide semantic analysis and causal inference.

Domain‑Graph Constraints: A dynamic task map that directs the LLM where to look, what to skip, and what to explore deeply, analogous to ControlNet guiding image generation. The graph defines cause/effect boundaries, start/end points, and supports iterative expansion with cause/effect miners, ensuring each inference is grounded and traceable.

End‑to‑End Decision Making: The system delivers not only analytical insights but also actionable, auditable risk decisions. It features deep domain perception (thousands of tokens per node), multi‑layer verification (summary → hallucination detection → multi‑model debate), full‑process transparency with visualized reasoning paths, and controllable automation with expert‑intervention hooks.

Summary and Outlook: While current token limits prevent full‑scale deployment, early results show strong capabilities in understanding complex heterogeneous graphs and multimodal knowledge. Ongoing research aims to scale the solution, further reduce hallucinations, and integrate tighter domain expertise, ultimately delivering a responsible, explainable AI‑driven risk‑control platform.