AutoProNE: Automated and Unsupervised graph representation learning.

Code for Automated Unsupervised Graph Representation Learning in TKDE'21.

Examples:

python main.py --emb ./emb/wiki_deepwalk.embedding --adj ./emb/wikipedia.ungraph --saved-path ./out/wiki_spectral.embedding

python main.py --emb ./emb/cora_dgi.embedding --dataset cora --concat-search --prop-types sc ppr heat gaussian

• --emb : path of input embedding
• --adj : path of edgelist format adjacency matrix.
• --concat-search : search the filters used to concat, default is True.
• --prop-types : types of filters to use, options : ['ppr', 'heat', 'gaussian', 'sc'].
• --max-evals : num of iterations of automl to optimize loss, default: 100.
• --loss : loss function used in AutoML searching, default: 'infomax'
• --no-eval: if set, do not evalute the embedding after propagation.
• --workers : the number of working threads in AutoML. default: 10. Try to set --workers to a larger number for faster training.
• --dataset : (optional).
• --saved-path : path to save embeddings.

Currently using optuna as AutoML tool.

Dataset

The datasets used in the paper could be downloaded from this link.