Ant Group's Self‑Developed Graph Neural Network Research: GeniePath and Bandit Sampler

In recent years, the rise of neural networks has driven data mining research, highlighting the importance of relational data for risk identification.

Graph neural networks (GNNs) map graph nodes to low‑dimensional embeddings, preserving similarity based on labels or structural proximity, using a message‑passing framework where nodes generate, propagate, and aggregate messages.

GNNs can express complex relational patterns that traditional neural networks or XGBoost cannot, such as predicting the size of connected subgraphs.

Ant Group’s research introduced GeniePath, an adaptive‑receptive‑path GNN published at AAAI 2019, which learns both breadth (neighbor importance) and depth (multi‑hop importance) adaptively, improving performance on benchmarks like protein‑protein interaction networks.

For GNN optimization, Ant Group proposed a bandit‑based sampler (NeurIPS 2020) that approximates the variance‑optimal sampling distribution without requiring full neighbor information, achieving near‑optimal variance and faster convergence.

Experimental results show GeniePath’s superior F1 scores and the bandit sampler’s robustness compared with existing methods.

References: Liu et al., “GeniePath: Graph Neural Networks with Adaptive Receptive Paths,” AAAI 2019; Liu et al., “Bandit Samplers for Training Graph Neural Networks,” NeurIPS 2020.