DP-GCN: Node Classification Based on Both Connectivity and Topology Structure Convolutions for Risky Seller Detection

Chen Zhe, Aixin SunarXiv 2021 · OpenReview

TL;DR

This paper introduces DP-GCN, a novel dual-path graph convolutional network that simultaneously leverages both node connectivity and local topology structure for more accurate node classification. By treating these two crucial aspects of network data separately and then intelligently fusing them, DP-GCN significantly improves performance on tasks like risky seller detection, outperforming existing methods.

What Problem Does This Paper Address?

In many real-world networks, like payment systems, understanding a node's role and behavior requires more than just knowing who it's directly connected to. Sellers in a payment network, for instance, have distinct business models (e.g., supplier, retailer) that manifest as specific local structural patterns (topology structures). Existing graph-based node classification methods, particularly graph convolutional networks (GCNs), primarily focus on connectivity (who is connected to whom) and often overlook these crucial local structural similarities. This limitation hinders their effectiveness in tasks where both types of information are vital, such as identifying fraudulent or "risky" sellers who might exhibit specific topological patterns in their transaction networks.

Methods

The core innovation is the DP-GCN (Dual-Path Graph Convolutional Network). It employs a dual-path architecture, processing connectivity and topology structure information in parallel:

  • C-GCN Module: This module handles the traditional graph convolution, capturing relationships based on direct node connectivity (edges in the graph).
  • T-GCN Module: This is where the novelty lies. Instead of directly applying convolution to raw topology, it introduces a role-based convolution. This approach first identifies "latent topology roles" (abstract representations of common structural patterns) and then propagates information between member nodes and these roles. This significantly reduces computational complexity while effectively capturing topology structure similarities.
  • Multi-Head Self-Attention Module: This module acts as a sophisticated fusion mechanism. It intelligently weighs and combines the outputs from both the C-GCN and T-GCN modules, allowing the model to learn which type of information is more important for the classification task at hand. This interactive alignment helps to overcome the limitations of simply concatenating features.

The learned unified embeddings from these modules are then fed into a classification layer for the final prediction.

Datasets

The paper evaluates DP-GCN on seven datasets:

  • Open Datasets:
    • Euro, Brazil, USA (airline networks, where airport activity level is the label)
    • BA (a labeled network for research)
    • Cora (a citation network)
  • Synthetic Datasets:
    • Acm1, Acm2 (generated for experimental purposes)
  • Real-World Benchmark:
    • Three large-scale payment networks from PayPal (SG-Jan, SG-Apr, SG-Jul) for risky seller detection. These datasets are characterized by their large size and imbalanced class distribution.

Key Insights and Ideas

  • Topology Matters: The research strongly emphasizes that local topology structure is as important as connectivity for many real-world node classification tasks, especially in domains like fraud detection.
  • Role-Based Convolution is Efficient: The proposed role-based convolution for topology structure is a significant contribution. It offers a computationally efficient way to model complex structural similarities without the quadratic cost of pairwise comparisons.
  • Attention for Fusion: Using multi-head self-attention to align and weigh information from different feature extractors (connectivity vs. topology) is crucial for achieving optimal performance. It allows the model to adaptively focus on the most relevant information.
  • Dual-Path Synergy: The interactive nature of the dual-path architecture, where both modules inform each other through the attention mechanism, leads to a more robust and comprehensive node representation.

Key Results

  • Superior Performance: DP-GCN consistently outperforms ten state-of-the-art baselines across all seven benchmark datasets, achieving significant improvements in accuracy and F-score (e.g., an average of >8% improvement on Euro and Brazil datasets).
  • Risky Seller Detection: In the PayPal case study, DP-GCN achieved the best performance in terms of Precision, Recall, and F-score for risky seller detection, demonstrating its practical applicability to large-scale, imbalanced real-world problems.
  • Ablation Studies: These studies confirmed that both the C-GCN and T-GCN modules are essential, and the multi-head attention mechanism significantly contributes to the model's effectiveness. Removing any of these components led to a noticeable performance drop.
  • Visualization: Node representation visualizations showed that DP-GCN can group nodes based on both connectivity and topology structure, unlike standard GCNs which primarily rely on connectivity.

Limitations and Future Work

  • Initialization of Topology Roles: While the paper explores different methods for initializing topology roles (e.g., using Struc2vec, Graphwave), the choice of initialization can impact performance. The paper notes that the results of different DP-GCN variants were comparable, suggesting robustness.
  • Computational Cost for Very Large Graphs: While role-based convolution improves efficiency, the paper acknowledges that for extremely large graphs, further optimizations might be needed.
  • Generalizability of Role-Based Convolution: The authors suggest that the role-based convolution is generalizable and can be extended to model other types of node properties beyond topology structure, potentially leading to multi-path convolution frameworks.
  • Real-World Enterprise Needs: The success on PayPal's payment networks highlights the model's alignment with real-world enterprise requirements, suggesting its potential for broader adoption in financial risk management and other sensitive applications.

Summary updated

Abstract

A payment network contains transactions between sellers and buyers. Detecting risky (or bad) sellers on such a payment network is crucial to payment service providers for risk management and legal compliance. In this research, we formulate this task as a node classification task. Specifically, we aim to predict a label for each seller in a payment network, by analysing its properties and/or interactions. Nodes residing in different parts of a payment network can have similar local topology structures. Such local topology structures reveal sellers' business roles, eg., supplier, drop-shipper, or retailer. We note that many existing solutions for graph-based node classification only consider node connectivity but not the similarity between node's local topology structure. Motivated by business need, we present a dual-path graph convolution network, named DP-GCN, for node classification. DP-GCN considers both node connectivity and topology structure similarity. The proposed model consists of three main modules: (i) a C-GCN module to capture connectivity relationships between nodes, (ii) a T-GCN module to capture topology structure similarities between nodes, and (iii) a multi-head self-attention module to align both properties. We evaluate DP-GCN on seven benchmark datasets against diverse baselines. We also provide a case study of running DP-GCN on three large-scale payment networks from PayPal, one of the leading payment service providers. Experimental results demonstrate DP-GCN's effectiveness and practicability.
BibTeX
@misc{zhe2021dp,
  author = {Chen Zhe and Aixin Sun},
  title = {{DP-GCN: Node Classification Based on Both Connectivity and Topology Structure Convolutions for Risky Seller Detection}},
  year = {2021},
  eprint = {2112.04757},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2112.04757},
}
Figure