Demonstration of REIGNN - REcommender Inductive Graph Neural Network | Version 1

Welcome to the official repo of the REIGNN demonstration app -- web service for scientific collaborations assessment powered by a GNN-based recommender system. Here we present the source code for the CIKM'22 paper "Demonstration of REIGNN: R&D Team Management with Graph Machine Learning".

Artyom Sosedka, Anastasia Martynova, Vladislav Tishin, Natalia Semenova, Vadim Porvatov

arXiv PDF: to be added.

Prerequisites

For the REIGNN model:

numpy==1.19.5
pandas==1.3.0
torch==1.8.1
torch-sparse==0.6.12  
torch-scatter==2.0.8
torch-cluster==1.5.9
torch-spline-conv==1.2.1
torch-geometric==2.0.3
wandb==0.12.9

Attention -- installing PyTorch Geometric for GPU can be tricky.

For backend:

python = "~3.8"
fastapi = "^0.73.0"
pydantic = { extras = ["dotenv"], version = "^1.8.2" }
uvicorn = "^0.17.4"
cookiecutter = "^1.7.3"
loguru = "^0.6.0"
pandas = "^1.4.0"
scipy = "^1.8.0"
networkx = "^2.6.3"
mypy = "^0.931"
ujson = "^5.1.0"

For frontend:

TBD

Dataset

We propose a unique dataset that could be used for the evaluation of the REIGNN model. The initial data was gathered from the Semantic Scholar Open Research Corpus and SCImago Journal & Country Rank website.

CS1021general Network type Co-authorship Citation Nodes 2798829 2504381 Edges 30796749 17934023 Clustering 0.714 0.138

CS1021demo Network type Co-authorship Citation Nodes 3808 12618 Edges 31332 34975 Clustering 0.644 0.199

In order to obtain the full dataset, it is required to download additional files via download.sh. The final revision of the file structure includes general and local parts of the dataset.

General part

We use the large subgraph CS1021_general extracted from Semantic Scholar Corpus as the basis for the CS1021_demo dataset. All papers belong to the period from January 1st, 2011 to December 31st, 2021, and related to the area of Computer Science.

• SSORC_CS_2010_2021_authors_edge_list.csv - common graph edge list.
• SSORC_CS_2010_2021_authors_edges_papers_indices.csv - common table describing relations between edges in a co-authorship graph (collaborations) and nodes in a citation graph (papers).
• SSORC_CS_2010_2021_authors_features.csv - table with one-hot encoded authors' research interests.
• SSORC_CS_2010_2021_papers_features_vectorized_compressed_32.csv - table with vectorized via Universal Sentence Encoder abstracts of papers.

Local part

• <...>authors.edgelist - edge list of a dataset citations graph.
• <...>papers.edgelist - edge list of a dataset co-authorship graph.
• <...>authors_edges_papers_indices.csv - table describing relations between edges in a co-authorship graph (collaborations) and nodes in a citation graph (papers).
• <...>papers_targets.csv - target values for each auxiliary task regarding edges in a co-authorship graph.

App running

In order to create .venv and install dependencies for the app running, you need to execute .backend/setup.sh. For server launch, it is required to use .backend/start.sh.

Model running

The multitask version of the REIGNN model is available at model/REIGNN.py. If you want to perform tests in a separate environment, you can use the following code:

import torch
import torch.nn as nn

from model.dataloader import get_data
from model.utils import run
from model.REIGNN import REIGNN

root_dir = '../'
dataset_name, split_name, split_number = 'CS1021small', '5_0.1', 0
citation_graph, train_data, val_data, test_data, authors_to_papers, batch_list_x, batch_list_owner = get_data(root_dir, dataset_name, split_name, split_number)

# Global
epochs_per_launch, lr = 15000, 0.001
device = 'cuda:0'

# Local
c_conv_num, c_latent_size, a_conv_num, a_latent_size = 2, 128, 3, 384
operator, link_size, heads = "hadamard", 128, 1 

# Multitask weights
mt_weights = [0.05, 0.05, 0.05, 0.05]

# W&B parameters
wandb_output, project_name, entity, group  = False, 'REIGNN', 'test_entity', 'test_group'

# define the model
model = REIGNN(citation_graph.to(device), heads, device,\
                            train_data.to(device), val_data.to(device), test_data.to(device),
                            authors_to_papers,
                            cit_layers = c_conv_num, latent_size_cit = c_latent_size,
                            auth_layers = a_conv_num, latent_size_auth = a_latent_size,
                            link_size = link_size).to(device) 

optimizer, criterion = torch.optim.Adam(model.parameters(), lr=lr), nn.L1Loss()
run(wandb_output, project_name, group, entity, mt_weights, model, optimizer, criterion, operator, batch_list_x, batch_list_owner, epochs_per_launch)

Contact us

If you have some questions about the code, you are welcome to open an issue or send me an email, I will respond to that as soon as possible.

License

Established code released as open-source software under the MIT license.

Citation

To be added