Source-linked AI summary
GraphFAS: A Distributed System for Automated Graph Feature Generation and Selection in Industrial Transaction Networks
Yice Luo, Yun Zhu, Xi Chen, Yongchao Liu, Xintan Zeng, Chengying Huan, Kai Zhang, Jinrui Zhang, Juelu Zhang, Jiajun Zheng
TL;DR
Industrial fraud detection needs graph-structured signals but remains constrained by costly expert feature engineering and GNN limitations in interpretability and deployment. GraphFAS combines parameter-free structural feature generation with distributed Boruta selection using median aggregation. Deployed at Alipay, it processes millions of seed nodes daily, achieves order-of-magnitude efficiency gains over manual engineering, and outperforms GNN-based approaches while retaining interpretable TreeSHAP-compatible features.
Problem
Industrial fraud detection relies on costly manual features that can miss relational patterns, while GNNs pose interpretability, latency, and semantic-validation challenges in financial risk control.
Method
GraphFAS generates explicit structural features through parameter-free multi-hop extraction and multi-scale aggregation, then applies distributed Boruta selection with median aggregation across partitions.
Results
GraphFAS outperforms GNN-based approaches while achieving order-of-magnitude efficiency gains over manual feature engineering in a two-year Alipay deployment.
Takeaways & Limitations
Decoupling feature generation from model training provides CPU-only, tabular-model-compatible features with native TreeSHAP explainability for large-scale industrial risk control.
Takeaways & Limitations
The industrial data is sampled and abstracted, does not represent real business situations at Ant Financial Services Group, and is not restorable to original data.
Abstract
from arXiv · showhide
Industrial fraud detection often relies on costly expert-crafted features that overlook graph-structured relational signals, while GNNs often do not meet the interpretability and deployment requirements of financial risk control. We propose GraphFAS (Graph Feature Automated Selection), a distributed feature selection procedure based on Boruta that bridges this gap through: (1) a non-parametric graph feature generation module that constructs explicit, interpretable structural features via multi-hop subgraph extraction and multi-scale aggregation without learned parameters; and (2) an automated distributed feature selection algorithm extending Boruta with median-based aggregation across partitions to robustly identify informative features at scale with minimal domain expertise. Compared with end-to-end GNN pipelines, GraphFAS decouples feature aggregation from model training, enabling direct integration with tabular models and direct compatibility with TreeSHAPbased explanations. Deployed in Alipay, GraphFAS delivers orderof-magnitude improvements in engineering efficiency while showing strong performance against expert-driven and graph-learning baselines on large-scale graphs.
1 Introduction
GraphFAS targets costly, manual graph feature engineering and the interpretability, latency, and semantic limitations of GNNs in financial risk control. It combines parameter-free structural feature generation with distributed Boruta-based selection for scalable, interpretable deployment.
- Motivation: Manual graph feature engineering can span over a month, requires specialized expertise, and may miss important topological patterns.The workflow includes expert analysis, large-scale simulation, and iterative stability evaluation.
- Motivation: GNNs create deployment challenges because opaque transformations hinder transparency, training adds latency, and embeddings lack explicit semantics for risk analysis.These constraints conflict with regulatory, low-latency, validation, and production requirements.
- Approach: GraphFAS generates explicit, interpretable structural features through multi-hop extraction and multi-scale aggregation without learned parameters.The resulting tabular features support TreeSHAP-compatible explanations and CPU-based distributed execution.
- Approach: Its distributed Boruta extension uses median aggregation across partitions to identify informative features robustly while reducing sensitivity to outlier partitions.The procedure reduces reliance on manual expert intervention for large-scale graph data.
- Evaluation: GraphFAS is evaluated on eight public benchmarks and three industrial datasets, with millions of seed nodes processed daily across risk-control scenarios.The system is deployed in Alipay and reports strong industrial-dataset results.
- Evaluation: The system achieves over 10× lower feature-engineering cycle time and substantially higher detection rates than expert-crafted baselines.These results are reported as part of the Alipay deployment evaluation.
2 Background and Related Work
GraphFAS reframes graph representation learning by separating non-parametric feature aggregation from downstream model training. It combines interpretable graph statistics with distributed Boruta-based selection to address scalability, auditability, and deployment constraints in industrial graph settings.
- Problem Formulation: GraphFAS decouples feature aggregation from end-to-end training, producing selected structural features for a lightweight predictive model.The pipeline computes graph statistics, applies automated selection, and then optimizes a task-specific predictor.
- Problem Formulation: Non-parametric aggregation computes graph-level and node-level statistics without gradient-based optimization, preserving explicit feature semantics.Graph metrics and neighbor aggregators integrate structural information while maintaining interpretability.
- Feature Selection: GraphFAS extends Boruta with distributed partition-wise importance estimation, cross-partition aggregation, and final tentative-feature ranking.The design targets feature-selection scalability while accounting for partition-level variation.
- Related Work: GNNs provide strong graph-learning performance but create transparency, latency, and semantic-validation challenges in industrial financial risk control.The related-work discussion contrasts opaque embeddings and end-to-end training with explicit tabular-compatible features.
- Related Work: GraphFAS generates inherently interpretable structural statistics compatible with native TreeSHAP explainability, unlike opaque graph embeddings and approximate post-hoc explanations.These features are intended for direct validation and use with tabular learning models.
- Research Gap and Motivation: GraphFAS addresses these gaps through distributed Boruta selection, non-parametric feature generation, and decoupled deployment without end-to-end retraining.The system is presented as a scalable alternative to manual engineering and graph-learning pipelines.
3 The GraphFAS Framework
GraphFAS generates interpretable multi-hop graph features without learned parameters, then selects informative features through a distributed Boruta-based procedure using median importance aggregation.
- Feature Generation: GraphFAS extracts 1-hop, 2-hop, and 3-hop ego-subgraphs to capture localized structural patterns at varying depths.
- Feature Generation: Graph metrics characterize connectivity, centrality, and neighborhood composition, combining intrinsic attributes with structural context.
- Feature Generation: Multi-scale aggregation uses Max, Min, Mean, Std, Sum, and Count to represent extreme behavior, distributions, cumulative effects, and density.
- Distributed Feature Selection: The selection stage compares candidate-feature importance with shuffled shadow features using a distributed Boruta-style procedure.
- Distributed Feature Selection: Median importance across partitions confirms features above all shadows, rejects features below all shadows, and ranks remaining tentative features.
- Explainability: Selected explicit structural statistics support direct TreeSHAP attribution without the approximation techniques required by GNN embeddings.
4 Distributed Implementation
GraphFAS uses a Manager-Worker implementation on GeaLearning, distributing graph operations and synchronizing feature importance across workers for scalable execution.
- System Architecture: GeaLearning separates control and computation through a Manager-Worker architecture with a Driver coordinating distributed execution.
- Data Distribution: The implementation partitions the dataset across worker processes using 1D Distributed Sample Transposition to balance computational load.
- Feature Generation: Workers generate k-hop subgraphs for assigned seed nodes in parallel, replicating shared edges to preserve subgraph completeness.
- Feature Selection: Each worker computes feature importance locally, after which the manager synchronizes and aggregates scores across partitions.
- Feature Selection: Median aggregation reduces sensitivity to outlier partitions caused by skewed class distributions.
5 Experiments
Experiments cover eight public benchmarks and three Alipay industrial transaction datasets, using dataset-appropriate splits, infrastructure, and evaluation metrics.
- Datasets: The evaluation includes eight public datasets from PyTorch Geometric and SNAP plus three industrial Alipay transaction networks.
- Datasets: Public benchmarks include citation, co-purchasing, and social networks, with 60%/20%/20% splits for small datasets and predefined splits for Flickr and Reddit.
- Datasets: Industrial datasets range from 2.8M to 285.4M edges, contain multi-relational transaction links, and have fewer than 0.1% positive samples.
- Implementation Settings: Public experiments use a 64-core CPU and 256GB RAM, while industrial experiments run on Kubernetes production clusters.
- Evaluation: Accuracy, Micro-F1, and AUC-ROC are used for small public, large public, and industrial datasets, respectively.
5.2 Public Graph Benchmarks
GraphFAS is compared with traditional machine learning, GNN, and feature-selection baselines on public graph benchmarks, achieving competitive or improved results while feature count affects performance.
- Baselines: The comparison includes twelve baselines spanning traditional ML, deep graph learning, and feature-selection methods.
- Comparison with GNN Methods: GraphFAS achieves competitive results with state-of-the-art GNNs on smaller datasets and over 1% absolute improvement on larger datasets.
- Comparison with Feature Selection Methods: GraphFAS achieves 10-20% absolute improvements over feature-selection baselines.
5.3 Industrial Case Studies
GraphFAS outperforms the industrial-strength GIPA baseline across all three industrial datasets, including under severe class imbalance and across two orders of magnitude in graph size.
- 6.14% absolute AUC improvement is achieved on Dataset1, with GraphFAS reaching 86.26% versus GIPA’s 80.12%.
- 99.90% AUC-ROC is achieved on Dataset3, compared with GIPA’s 95.21% on a graph containing 285.37M edges.
- GraphFAS maintains strong performance under severe class imbalance, with fewer than 0.1% positive samples.
- Consistent improvements across graph sizes from approximately 3M to approximately 285M edges suggest good distributed scalability.
5.4 Efficiency Analysis and Ablation Study
GraphFAS improves distributed execution as containers increase and achieves a favorable performance–runtime trade-off through graph feature generation and feature selection.
- Efficiency Analysis: 3.6× overall speedup is achieved when scaling Dataset3-97M from 16 to 64 containers, although sublinear scaling reflects growing communication and synchronization overheads.
- Efficiency Analysis: 500 features are selected across datasets because performance stabilizes around that count, while additional features can introduce noise.
- Ablation Study: The full model provides the best overall trade-off between performance and runtime.
- Ablation Study: The full model significantly outperforms the version without graph feature generation across all datasets.
- Ablation Study: On Computers, GraphFAS reaches 91.54% accuracy versus 90.89% without feature selection, while running over 10× faster.
6 Deployment and Industrial Impact
GraphFAS has operated in production for over two years, using an expert-supervised workflow to generate, validate, and deploy graph features for real-time risk control. In a cash-out fraud scenario, it increased latent fraud-group discovery and shortened feature engineering substantially.
- GraphFAS has been deployed for over two years, handling millions of seed nodes daily across multiple risk control scenarios.
- The operational workflow proceeds from seed selection and offline graph mining to expert validation and registration for real-time inference.
- A tenfold lift in uncovering latent fraud groups was achieved in a representative cash-out fraud detection scenario compared with baseline methods.
- Automating feature discovery reduced the feature engineering cycle by over 10× compared with traditional manual assessment.
7 Conclusion
GraphFAS combines non-parametric graph feature generation with distributed Boruta-style selection to provide interpretable, scalable features for industrial risk control. Its deployment demonstrates practical efficiency and explainability benefits, while accepting some representational-capacity trade-off.
- GraphFAS combines non-parametric graph feature generation with median-aggregated distributed Boruta-style selection.
- The system produces explicit structural features compatible with native TreeSHAP explainability and supports CPU-only distributed processing.
- After two years of Alipay deployment, GraphFAS processes millions of seed nodes daily and achieves order-of-magnitude efficiency gains over manual feature engineering.
- Decoupling feature generation from model training sacrifices some representational capacity in exchange for audit-compliant statistics and flexible downstream model updates.
8 GenAI Usage Disclosure
The authors disclose using Claude Code for code development and manuscript writing, with all AI-generated content reviewed, verified, and refined by the authors.
- Claude Code assisted with boilerplate code, implementation details, drafting and polishing text, and language clarity.