SIGEM: A Simple yet Effective Similarity based Graph Embedding Method (Best Research Paper, Honorable Mention at KDD'25)

This repository provides the reference implementation of SIGEM and LINOW (linear matrix form, LINOW-sn, LINOW-bn/MP, and LINOW-bn/TF).

SIGEM is a self-supervised and contrastive-free embedding method that provides a single latent vector for each node in a graph. SIGEM is simply implemented by a single-layer neural network with a listwise learning-to-rank loss function:

(1) SIGEM employs LINOW, as a link-based similarity measure, to compute similarity scores of nodes batch-wise.

(2) For each nodes $u$ in the batch, it then ranks the nodes based on their LINOW scores w.r.t. $u$ in descending order and picks the top-t ones.

(3) After completing the ranking process for all nodes in the graph, SIGEM tries to preserve the original ranks of the top-t similar nodes to any node $u$, within their corresponding latent vectors in the embedding space.

Installation and Usage

In order to run SIGEM, the following packages are required:

Python       = 3.12
tensorflow   = 2.18.0
numpy        = 1.26.4
scipy        = 1.13.1
scikit-learn = 1.5.2
tqdm         = 4.66.6

SIGEM can be run directly from the command line or migrated to your favorite IDE.

Graph Format

1. A graph must be represented as a text file under the edge list format in which, each line corresponds to an edge in the graph, tab is used as the separator of the two nodes, and the node index is started from 0.
2. A single original link in an undirected graph and its train/test splitted graphs must be represented via two links in both directions.

Running SIGEM

SIGEM has the following parameters:

--graph: Input graph; default is ''

--dataset_name: dataset name; default is None

--result_dir: Destination to save the embedding result, default is "output/" in the root directory

--dim: The embedding dimension, default is 128

--itr: Number of Iterations to compute LINOW, default is 5

--damping_factor: Damping factor for similarity computation, defaul is 0.2

--scaling_factor: Scaling factor to select top nodes (defaul is None); set it as 10 for link prediction and node classification tasks, and as 2 for the graph reconstruction task

--gpu: The flag indicating to run SIGEM on GPU, default is True

--bch_cpu: Batch size for computation on CPU; default is 128. It is ignored if --gpu=True

--bch_gpu: Batch size for computation on GPU; default is 128. It is ignored if --gpu=False

--prl_num: Number of parallel computation on CPU; default is 8. It is ignored if --gpu=True

--epc: Number of Epochs for training, default is 100

--lr: Learning rate (defaul is None); set it as 0.0030 for small graphs and as 0.0012 for very large graphs

--early_stop: Flag indicating to stop the training process if the loss stops improving, default is True

--wait_thr: Number of epochs with no loss improvement after which the training will be stopped, default is 10

--read_topK_nodes: Flag to read top nodes from a binary file, they will be selected by computing similarity otherwise; default is False

--topK_file: Paths to the saved top nodes

--write_topK_nodes: Flag to write top nodes into a binary file for later usage; default is False

--topK_save_path: Path to write the top nodes into a binary file

NOTE:

It is highly recommended to set the values of (1) "--itr" and "--damping_factor" to 5 and 0.2, respectively, for all tasks, (2) "--scaling_factor" to 10, 10, and 2 for link prediction, node classification, and graph reconstruction tasks, respectively, (3) "--bch_cpu", "--bch_gpu", and "--lr" parameters to the ones explained in Section 5.3.1 of the paper.

Sample:

1. Link Prediction Task

python SIGEM.py --graph data/Cora_train_70.txt --dataset_name Cora_train_70 --scaling_factor 10 --lr 0.0030

python SIGEM.py --graph data/Cora_train_70.txt --dataset_name Cora_train_70 --scaling_factor 10 --lr 0.0030 --write_topK_nodes True --topK_save_path output/ ## saving top nodes in a file before training

python SIGEM.py --graph data/Cora_train_70.txt --dataset_name Cora_train_70 --scaling_factor 10 --lr 0.0030 --read_topK_nodes True --topK_file output/Live_train_70_SIGEM_IT_5_C_2_listMLE_topK_Scl10 ## loading top nodes from a file for training

2. Node Classification Task

python SIGEM.py --graph data/Cora_directed_graph.txt --dataset_name Cora --scaling_factor 10 --lr 0.0030

3. Graph Reconstruction Task

python SIGEM.py --graph data/Live_undirected_graph.txt --dataset_name Live --scaling_factor 2 --lr 0.0030

Citation:

Masoud Reyhani Hamedani, Jeong-Seok Oh, Seong-Un Cho, and Sang-Wook Kim. 2025. SIGEM: A Simple yet Effective Similarity based Graph Embedding Method. In Proceedings of the 31st ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD'25, August 03-07, 2025, Toronto, ON, Canada, pages: 2420-2431, https://doi.org/10.1145/3711896.3737128