sonaidgr8/uss_nmf

A Unified Non-Negative Matrix Factorization Framework for Semi Supervised Learning on Graphs

Semi-supervised clusterable node representation learning for homogeneous graphs.

Please see our initial work:~ Vijayan, Priyesh, et al. Semi-supervised learning for clusterable graph embeddings with NMF, Poster:~ NeuIPS 2018 Relational learning Workshop

How to run

Input:-

Input is organized as follows -
Datasets/
    |_ _ _ <Dataset-name>/
                |_ _ _ <Dataset-name.mat>
                |_ _ _ <Stats.txt>
                |_ _ _ <Percentage of train-test splits>/
                                    |_ _ _ <Fold-Number>/
                                                |_ _ _ train_ids.npy
                                                |_ _ _ test_ids.npy
                                                |_ _ _ val_ids.npy

Usage:-

python main_algo.py --DATA_DIR cora --ALPHA 1 --BETA 0.1 --THETA 0.5 --K 20 --L_COMPONENTS 128
    * ALPHA : Similarity matrix factorization weight
    * BETA : Community Indicator matrix factorization weight
    * THETA : Label matrix factorization weight
    * K : Number of clusters
    * L_COMPONENTS : Dimension of representation space

Please look at the get_ArgumentParser() function in main_algo.py to specify default values.

Output:-

1. The generated node and label embeddings are saved in emb/ folder as Q & U .npy files respectively.
2. The node and label embeddings are of dimension (#Nodes x L_COMPONENTS) & (#Labels x L_COMPONENTS).
3. The Node Classification evaluation results are stored in Results folder.

License

This project is licensed under the MIT License - see the LICENSE.md file for details