Ant Financial’s KDD 2018 Papers: Graph-Based Fraud Detection, GeniePath GNN, and Distributed Collaborative Hashing

At the 2018 ACM SIGKDD Conference in London, Ant Financial contributed three papers that showcase the application of advanced graph‑learning and distributed machine‑learning techniques to real‑world problems.

Paper 1: "Who‑Stole‑the‑Postage? Fraud Detection in Return‑Freight Insurance Claims" addresses large‑scale insurance fraud by constructing account‑relationship graphs (transfer graphs and device‑sharing graphs) and applying Graph Neural Networks (GNNs), specifically the GeniePath algorithm, to identify fraudulent accounts. The study compares GeniePath with node2vec and GBDT, reporting a >35% increase in captured black‑market samples and a 20%+ improvement in F1 score.

Paper 2: "GeniePath: Graph Neural Networks with Adaptive Receptive Paths" introduces an adaptive GNN that learns to select valuable neighbors during aggregation, using attention for breadth and an LSTM‑style module for depth. The paper details the mathematical formulation, iterative update rules, and demonstrates superior perception‑field selection compared with standard GCNs on benchmark protein‑network datasets, achieving up to 20% higher F1.

Paper 3: "Distributed Collaborative Hashing and Its Applications in Ant Financial" proposes a hash‑based matrix factorization framework that leverages a parameter‑server architecture for scalable training. By relaxing binary hash vectors to real‑valued embeddings, performing SGD updates, and then binarizing, the method dramatically reduces offline training time and online scoring latency while maintaining comparable recommendation accuracy on Netflix and internal Alipay data.

All three works illustrate how Ant Financial integrates cutting‑edge AI research—graph neural networks, adaptive neighbor selection, and distributed hashing—into its risk control and recommendation pipelines, delivering faster, more accurate, and scalable solutions for billions of users.