🌐 Documentation | 🛠 Development
The l2gv2 package is a framework for scalable graph
representation learning, implementing algorithms such as the one depicted in
Jeub, L.G.S., Colavizza, G., Dong, X., Bazzi, M., Cucuringu, M.
Local2Global: a distributed approach for scaling representation learning on graphs. Mach Learn 112, 1663–1692 (2023). https://doi.org/10.1007/s10994-022-06285-7
Traditional centralized graph representation learning methods do not scale well to large graphs, which do not fit into memory. The core idea of local2global algorithms (both the one presented above, and a new algorithm developed in this package) is to divide a large graph into smaller, overlapping subgraphs -- referred to as patches and learn local embeddings independently. These local embeddings are then aligned by learning appropriate transformations between patches so that the patches can be made to overlap in a shared global embedding. Alignment takes place by inferring appropriate transformation matrices for rescaling, rotation and translation as shown in the figure above, adapted from the Local2Global paper.
This library implements local2global algorithms, allowing patches to be processed independently. The library includes methods for graph partitioning, local embedding generation using popular graph neural network (GNN) models such as Graph AutoEncoders (GAE), and algorithms for global alignment. We also show the application of the library in correctly estimating global embeddings for well-known graph neural network models in the literature, such as the Cora dataset of machine learning citations and the DGraph financial dataset.
The package is available on PyPI at TODO
For detailed installation instructions, consult the documentation.
The following people contributed to this project (emoji key).
This project follows the all-contributors specification.