Source-linked AI summary
A Semi-supervised Graph Attentive Network for Financial Fraud Detection
Daixin Wang, Jianbin Lin, Peng Cui, Quanhui Jia, Zhen Wang, Yanming Fang, Quan Yu, Jun Zhou, Shuang Yang, Yuan Qi
TL;DR
Financial fraud detection is challenged by heterogeneous user interactions and scarce labels, which conventional rule-based and feature-based methods do not fully exploit. The paper proposes SemiGNN, a semi-supervised multiview graph neural network with hierarchical attention and interpretable outputs. On Alipay data, it reports better accuracy than state-of-the-art methods on two tasks and provides task-related interpretive insights.
Problem
Fraud detection must use heterogeneous user interactions and multiview information despite only very few users being labeled.
Method
SemiGNN expands labeled users through social relations and applies hierarchical attention to jointly model labeled and unlabeled multiview graph data.
Results
The method achieves better accuracy than state-of-the-art methods on two Alipay fraud-detection tasks and produces interpretable results.
Takeaways & Limitations
Attention-based modeling identifies important factors for a task and offers insights into why users are predicted as fraud.
Abstract
from arXiv · showhide
With the rapid growth of financial services, fraud detection has been a very important problem to guarantee a healthy environment for both users and providers. Conventional solutions for fraud detection mainly use some rule-based methods or distract some features manually to perform prediction. However, in financial services, users have rich interactions and they themselves always show multifaceted information. These data form a large multiview network, which is not fully exploited by conventional methods. Additionally, among the network, only very few of the users are labelled, which also poses a great challenge for only utilizing labeled data to achieve a satisfied performance on fraud detection. To address the problem, we expand the labeled data through their social relations to get the unlabeled data and propose a semi-supervised attentive graph neural network, namedSemiGNN to utilize the multi-view labeled and unlabeled data for fraud detection. Moreover, we propose a hierarchical attention mechanism to better correlate different neighbors and different views. Simultaneously, the attention mechanism can make the model interpretable and tell what are the important factors for the fraud and why the users are predicted as fraud. Experimentally, we conduct the prediction task on the users of Alipay, one of the largest third-party online and offline cashless payment platform serving more than 4 hundreds of million users in China. By utilizing the social relations and the user attributes, our method can achieve a better accuracy compared with the state-of-the-art methods on two tasks. Moreover, the interpretable results also give interesting intuitions regarding the tasks.
I. INTRODUCTION
Financial fraud detection is difficult because conventional methods rely on rules or individual statistical features, while financial users generate heterogeneous relational data and only a small fraction are labeled. The paper proposes SemiGNN, a semi-supervised multiview graph-attention model, and reports gains over state-of-the-art methods on two Alipay tasks.
- Financial fraud harms users and service providers, making fraud detection an important research problem.
- Rule-based methods depend on human knowledge, struggle with changing complex patterns, and can be attacked.
- Existing machine-learning methods mainly use statistical features from individual users, overlooking their interactions in financial networks.
- The paper identifies three challenges: bridging scarce labeled data with unlabeled users, integrating heterogeneous views, and learning an interpretable model.
- SemiGNN links labeled and unlabeled users through social relations, models relational and attribute data as a multiview network, and applies hierarchical attention.
- Experiments on Alipay users and two tasks report a substantial gain over state-of-the-art methods, alongside interpretable results.
B. Learning over graphs
The paper represents multiple facets of user information as graph views, including relational and attribute graphs, and trains a classifier over labeled and socially reached unlabeled users. Its graph formulation supports modeling heterogeneous user information within a multiview network.
- Learning over graphs: Network embedding learns low-dimensional node representations, while graph-based methods extend this modeling to topology and node features.
- Learning over graphs: Graph-based semi-supervised learning uses labeled and unlabeled vertices, encouraging label consistency and similar predictions for similar vertices.
- Problem Definition and Notations: The problem collects multiple views of user information, including social or transaction relations and attribute graphs constructed instead of dense features.
- Problem Definition and Notations: The notation defines nL labeled users and nuL socially reached unlabeled users, with nuL ≫ nL, across m view-specific graphs.
- Problem Definition and Notations: View-specific graphs may contain users plus apps or attribute nodes, with edges representing the corresponding user relationships or attributes.
B. SemiGNN
SemiGNN uses hierarchical attention to integrate neighbors within each graph view and combine heterogeneous multiview information for user representation.
- SemiGNN answers three design questions by assembling neighbors, integrating views, and jointly modeling labeled and unlabeled users.
- Node-level Attention: Node-level attention learns neighbor importance within each view and aggregates neighbor representations into low-level view-specific user embeddings.
- Node-level Attention: For an edge-weighted neighbor, the model multiplies the edge weight by the neighbor’s embedding before aggregation.
- Node-level Attention: The framework separately models each view’s graph structure and obtains corresponding user embeddings for other views similarly.
3) View-level Attention:
SemiGNN projects view-specific embeddings into a shared high-level space before using view-level attention to fuse their complementary information.
- Separate MLPs project low-level view-specific user embeddings into high-level space before multiview integration.
- The view-level attention mechanism learns the importance of different views and integrates them according to their task relevance.
- A user-specific preference vector guides view attention, assigning greater importance when it is similar to a view-specific vector.
- The joint user embedding is formed by weighted combination and concatenation of view-specific embeddings.
- A one-layer perceptron refines the joint embedding into the final high-level user representation for task-specific loss computation.
4) Loss function and Optimization:
SemiGNN combines supervised classification with an unsupervised graph objective so unlabeled multiview users contribute to representation learning.
- Supervised Loss: For labeled users, softmax over embedding representations produces classification results and defines the supervised classification loss.
- Only a small portion of users are labeled, motivating use of their available multiview information during training.
- Unsupervised Graph Loss: Random walks define graph neighbors, and the unsupervised loss makes nearby representations similar while separating disparate nodes.
- Unsupervised Graph Loss: The graph loss uses positive co-occurring neighbors, negative neighbors, a negative-sampling distribution, and Q=3 negative samples.
- Objective and Optimization: Unlabeled representations are generated from multiview information, allowing the model to use both social relations and content information.
- Objective and Optimization: LSemiGNN = α · Lsup + (1 − α) · Lgraph + λLreg combines supervised and graph losses with L2 regularization.
5) Analysis of the Proposed Model:
The analysis presents SemiGNN as an efficient, general, interpretable, and inductive multiview graph model that can classify new users with multiview information.
- SemiGNN is designed as a general multiview graph model for various node types and graph views.
- Its unified optimization fuses view semantics and allows different views to mutually promote one another.
- Node-level and view-level attention can identify influential nodes and views associated with a user’s fraud classification.
- Given a new user’s multiview information, the inductive model can directly classify the user.
- Training complexity is O(I · |E| · d · m), and the model’s complexity is linear in the relation-graph edge count.
IV. EXPERIMENT
The experiments use ALIPAY data to evaluate SemiGNN on default prediction and user-attribute prediction, including questions about embeddings, attention, unlabeled data, and parameter sensitivity.
- Research Questions: The experiments test comparative embedding quality, hierarchical attention, benefits from unlabeled data, and parameter sensitivity.These questions are evaluated on a real-world dataset.
- Experimental Tasks: The study evaluates SemiGNN on ALIPAY, a large Chinese online and offline payment platform, using default and user-attribute prediction tasks.The attribute task predicts user occupation.
- Data Construction: The multiview graph combines user relations, user-app interactions weighted by login frequency, and user-nick connections.These sources represent social, behavioral, and attribute-related information.
- Data Construction: About 4 million users have known labels, while one-hop friends, classmates, and workmates expand the dataset to over 100 million users.Most users have app, nick, and address information, supporting multiview modeling.
B. Baseline Methods
The evaluation compares SemiGNN with conventional classifiers, graph representation methods, graph neural networks, and reduced variants that remove supervision or attention components.
- Conventional and Representation Baselines: Xgboost uses dense features from each view as input to a tree-based classifier.
- Conventional and Representation Baselines: LINE learns user representations on the multiview graph before softmax classification and is reported as a scalable representative of related graph embedding methods.
- Graph Neural Baselines: GCN ensembles each node’s embedding with neighbor embeddings using labeled users as samples and averaged pretrained attribute embeddings as node attributes.
- Graph Neural Baselines: GAT extends GCN with attention-based neighborhood aggregation.
- SemiGNN Variants: SemiGNNsup uses only labeled data, while SemiGNNnd and SemiGNNvw remove node-level and view-level attention, respectively.
- Training Setup: The reported implementation trains SemiGNN with Adam for 3 epochs and averages results across 3 repeated experiments.
1) User Default Prediction:
SemiGNN and its variants outperform the compared methods on default and attribute-related evaluations. The results also support integrating multiple views, both attention levels, and unlabeled relational data.
- Default Prediction: SemiGNN and its variants perform better than other methods on the default-prediction evaluation.Default prediction uses AUC and KS as evaluation metrics.
- Default Prediction: SemiGNN outperforms SemiGNNnd and SemiGNNvw, indicating that node-level and view-level attention are both essential in the model.
- Default Prediction: SemiGNNsup outperforms GAT and GCN, supporting the use of differentiated attention across neighbors and views.The attention mechanism also provides interpretable results.
- Default Prediction: The relatively large improvement from SemiGNNsup to SemiGNN indicates that unlabeled users contribute valuable social-relational and multiview information.
- Attribute Prediction: SemiGNN and its variants outperform other methods on the three attribute-prediction metrics: F1-score, precision, and recall.The occupation task also evaluates top-ranked precision for financial risk control.
- View Effects: Using only one graph view substantially reduces performance on both tasks compared with SemiGNN.The result supports integrating multiple views for more comprehensive prediction.
- View Effects: Social relations are effective in both tasks, while apps and addresses are especially important for default prediction.Addresses can indirectly reflect users’ financial condition, and app use is linked to borrowing and spending.
- View Effects: Addresses are most important for occupation prediction, whereas app features perform worse because few apps are specific to particular occupations.
E. Interpretable Results
SemiGNN uses node-level attention to identify important neighbors and aggregates these weights into global feature importance. The reported patterns connect influential apps, nicks, and addresses to the prediction tasks.
- Attention-Based Interpretation: Node-level attention weights represent neighbor importance for a user, and aggregating them yields global importance for the task.
- Default Prediction: For default prediction, important apps are mostly financial, especially peer-to-peer lending apps, with game and shopping apps also influential.The paper relates these app categories to borrowing or spending that may reduce repayment capacity.
- Reported Features: The interpretation table reports the 15 most important words for default and attribute prediction, translating Chinese terms and categorizing apps.
- Occupation Prediction: For doctor-occupation prediction, important nicks often contain occupation titles, while important addresses include hospital departments.These patterns provide task-specific textual and location-related signals.
F. Parameter Sensitivity
Parameter sensitivity experiments on ALIPAY default prediction show that final embedding size and the balance between labeled and unsupervised information materially affect AUC, whereas initial embedding size has little effect.
- Final embedding dimension: AUC rises and then drops as the final embedding dimension grows.A suitable dimension encodes enough information, while larger dimensions can cause overfitting and redundancy.
- Initial node embedding dimension: Changing the initial node embedding dimension generally produces little AUC change.The tested dimensions are sufficient to encode useful information.
- Value of α: Performance improves when the model pays more attention to labeled data, because labels provide discriminative information.Ignoring unsupervised information nevertheless worsens performance because it provides structural information beneficial to learning.
- Model context: The model links labeled and unlabeled data through social relations and integrates multiview information using hierarchical attention.Node-level attention correlates neighbors, while view-level attention integrates different views.
- Future work: Future work may differentiate different social relations to further improve the model.The paper also suggests extending the model to additional fraud-detection applications.