IEEE ICDM 2023 Graph Learning Challenge: Community Detection and Fraud Group Mining

Recently the IEEE ICDM 2023 Graph Learning Challenge finals were held in Shanghai, jointly organized by Ant Group and Zhejiang University. The four‑month competition used a three‑stage format (preliminary, semi‑final, final) and attracted six teams, including participants from NetEase and Huazhong University of Science and Technology.

IEEE ICDM (International Conference on Data Mining) is one of the three top conferences in data mining, alongside ACM SIGKDD and SIAM SDM, covering algorithms, software, systems, and applications across data mining and artificial intelligence.

The competition focused on the “community detection” problem, aiming to partition a graph into densely connected sub‑communities to reveal hidden structures. Community detection is widely applied in social network analysis, bio‑informatics, risk control, recommendation, and especially in fraud detection where grouping similar users aids black‑market tracking.

With the rapid growth of online networks, community detection and fraud‑group mining have become increasingly complex. Deep graph learning, particularly pretrained models, can automatically learn high‑level representations to improve detection performance.

Teams accessed the Ant TuGraph platform to download the ICDM2023 pretrained and test datasets, both homogeneous graphs. The test set contains only topology and node attributes without community labels, and its vertex IDs do not correspond to those in the training set, making label prediction a challenging task.

The winning team from NetEase Yidun presented a novel fraud‑group detection algorithm called Risk‑DCRN. They first performed community pre‑partitioning, then applied a Dual Correlation Reduction Network (DCRN) for dense sub‑graph self‑supervised clustering, ultimately identifying 346 fraudulent groups in a large‑scale relational network under sparse label conditions.

Judges praised the competition for encouraging innovative use of pretrained models in real‑world scenarios, noting the high quality of technical reports and presentations and declaring the event a complete success.