Graph Neural Networks for Friend Ranking in Large-scale Social Platforms

Contributors: Aravind Sankar (asankar3@illinois.edu).

Aravind Sankar, Yozen Liu, Jun Yu, and Neil Shah, "Graph Neural Networks for Friend Ranking in Large-scale Social Platforms", The Web Conference 2021, WWW 2021, April 19–23, 2021, Ljubljana, Slovenia.

This repository contains a PyTorch implementation of GraFrank - a graph neural network architecture for multi-faceted friend ranking with multi-modal node features and pairwise link features in social networks. GraFrank uses two modules: modality-specific neighbor aggregation and cross-modality attention to inductively compute user representations in a dynamic friendship graph.

Since we are unable to publicly share the Snapchat datasets and industry-scale implementation of GraFrank (privacy reasons), we implement the core modules of GraFank using a popular geometric deep learning libray PyTorch Geometric.

Requirements

The code has been tested running under Python 3.6 using PyTorch 1.8.1 and PyTorch Geometric 1.7.0 (along with their dependencies).

To guarantee that you have the right package versions, you can use Anaconda to set up a virtual environment and following the instructions here.

Input Format

This repository uses a sample dataset based on the Cora citation network. To ensure compatibility with link features, we use dummy features in the code. Note that Cora is not meant to reproduce our experimental results.

In order to use your own data, you should use the data format of PyTorch Geometric and may need to update the data loader if the graph is dynamic. The code assumes that the node features are first partitioned and then concatenated across the K modalities. The relevant feature dimensions for the modality partitions are provided during model initialization.

Repository Organization

• models/ contains:

  • GraFRank model (GraFRank.py);
  • GraphSAGE model (SAGE.py);
  • modality-specific aggregators and cross-modality attention modules in GraFrank (layers.py)

• utils/ contains:

  • mini-batch loader for sampling positive and negative examples for model training (sampler.py);

Running the Model

To train and evaluate the model, please run

python train.py

Reference

If you make use of this code or the GraFrank algorithm in your work, please cite the following paper:

@inproceedings{sankar2021graph,
  title     = {Graph Neural Networks for Friend Ranking in Large-scale Social Platforms},
  author={Sankar, Aravind and Liu, Yozen and Yu, Jun and Shah, Neil},
  doi       = {10.1145/3442381.3450120},
  booktitle = {Proceedings of The Web Conference 2021},
  year = {2021}
}