G-Meta: Graph Meta Learning via Local Subgraphs

Authors: Kexin Huang, Marinka Zitnik

Project Website

Prevailing methods for graphs require abundant label and edge information for learning. When data for a new task are scarce, meta learning can learn from prior experiences and form much-needed inductive biases for fast adaption to new tasks.

Here, we introduce G-Meta, a novel meta-learning algorithm for graphs. G-Meta uses local subgraphs to transfer subgraph-specific information and learn transferable knowledge faster via meta gradients. G-Meta learns how to quickly adapt to a new task using only a handful of nodes or edges in the new task and does so by learning from data points in other graphs or related, albeit disjoint label sets. G-Meta is theoretically justified as we show that the evidence for a prediction can be found in the local subgraph surrounding the target node or edge.

Experiments on seven datasets and nine baseline methods show that G-Meta outperforms existing methods by up to 16.3%. Unlike previous methods, G-Meta successfully learns in challenging, few-shot learning settings that require generalization to completely new graphs and never-before-seen labels. Finally, G-Meta scales to large graphs, which we demonstrate on a new Tree-of-Life dataset comprising of 1,840 graphs, a two-orders of magnitude increase in the number of graphs used in prior work.

Environment Installation

python -m pip install --user virtualenv
python -m venv gmeta_env
source activate gmeta_env
pip install -r requirements.txt

Run

cd G-Meta
# Single graph disjoint label, node classification (e.g. arxiv-ogbn)
python train.py --data_dir DATA_PATH --task_setup Disjoint
# Multiple graph shared label, node classification (e.g. Tissue-PPI)
python train.py --data_dir DATA_PATH --task_setup Shared
# Multiple graph disjoint label, node classification (e.g. Fold-PPI)
python train.py --data_dir DATA_PATH --task_setup Disjoint
# Multiple graph shared label, link prediction (e.g. FirstMM-DB, Tree-of-Life)
python train.py --data_dir DATA_PATH --task_setup Shared --link_pred_mode True

It also supports various parameters input:

python train.py --data_dir # str: data path
                --task_setup # 'Disjoint' or 'Shared': task setup, disjoint label or shared label
                --link_pred_mode # 'True' or 'False': link prediction or node classification
                --batchsz # int: number of tasks in total
                --epoch # int: epoch size
                --h # 1 or 2 or 3: use h-hops neighbor as the subgraph.
                --hidden_dim # int: hidden dim size of GNN
                --input_dim # int: input dim size of GNN
                --k_qry # int: number of query shots for each task
                --k_spt # int: number of support shots for each task
                --n_way # int: number of ways (size of the label set)
                --meta_lr # float: outer loop learning rate
                --update_lr # float: inner loop learning rate
                --update_step # int: inner loop update steps during training
                --update_step_test # int: inner loop update steps during finetuning
                --task_num # int: number of tasks for each meta-set
                --sample_nodes # int: when subgraph size is above this threshold, it samples this number of nodes from the subgraph
                --task_mode # 'True' or 'False': this is specifically for Tissue-PPI, where there are 10 tasks to evaluate.
                --num_worker # int: number of workers to process the dataloader. default 0.
                --train_result_report_steps # int: number to print the training accuracy.

To apply it to the five datasets reported in the paper, using the following code as example after you download the processed datasets from the section below.

arxiv-ogbn:

Tissue-PPI:

Fold-PPI:

FirstMM-DB:

Tree-of-Life:

Also, check out the Jupyter notebook example.

Data Processing

We provide the processed data files for five real-world datasets in this Drive folder and this Microsoft OneDrive folder.

1) To create your own dataset, create the following files and organize them as follows:

• graph_dgl.pkl: A list of DGL graph objects. For single graph G, use [G].
• features.npy: An array of arrays [feat_1, feat_2, ...] where feat_i is the feature matrix of graph i.

2.1) Then, for node classification, include the following files:

• train.csv, val.csv, and test.csv: Each file has two columns, the first one is 'X_Y' (node Y from graph X) and its label 'Z'. Each file corresponds to the meta-train, meta-val, meta-test set.
• label.pkl: A dictionary of labels where {'X_Y': Z} means the node Y in graph X has label Z.

2.2) Or, for link prediction, note that the support set contains only edges in the highly incomplete graph (e.g., 30% of links) whereas the query set edges are in the rest of the graph (e.g., 70% of links). In the neural message passing, the GNN should ONLY exchange neural messages on the support set graph. Otherwise, the query set performance is biased. Because of that, we split the meta-train/val/test files into separate support and query files. For link prediction, create the following files:

• train_spt.csv, val_spt.csv, and test_spt.csv: Two columns, first one is 'A_B_C' (node B and C from graph A) and the second one is the label. This is for the node pairs in the support set, i.e. positive links should be in the underlying GNN graph.
• train_qry.csv, val_qry.csv, and test_qry.csv:Two columns, first one is 'A_B_C' (node B and C from graph A) and the second one is the label. This is for the node pairs in the query set, i.e. positive links should NOT be in the underlying GNN graph.
• train.csv, val.csv, and test.csv: Merge the above two csv files.
• label.pkl: A dictionary of labels where {'A_B_C': D} means the node B and node C in graph A has link status D. D can be 0 or 1 means no link or has link.

We also provide a sample data processing scripts in data_process folder. See node_process.py and link_process.py.

Cite Us

@article{g-meta,
  title={Graph Meta Learning via Local Subgraphs},
  author={Huang, Kexin and Zitnik, Marinka},
  journal={NeurIPS},
  year={2020}
}

Contact

Open an issue or send an email to kexinhuang@hsph.harvard.edu if you have any question.