keyword

scalable self-supervised graph clustering

Scalable self-supervised graph clustering is an unsupervised machine learning approach designed to partition nodes or subgraphs within massive graph-structured datasets into cohesive clusters without relying on manual labels. It combines graph representation learning with self-supervised pretext objectives, such as contrastive learning, data augmentation discrimination, and structural reconstruction, to capture both node attributes and topological connectivity. To overcome the prohibitive memory and computational bottlenecks typically encountered when applying full-graph neural networks to datasets with millions or billions of elements, it utilizes scalable training strategies such as mini-batch optimization, neighborhood sampling, or decoupled feature propagation. This design allows representation learning and cluster distribution optimization to proceed simultaneously and efficiently, enabling the discovery of natural community structures across large-scale networks.

1 item