Vertical Federated Graph Neural Network for Recommender System

Authors: Peihua Mai, Yan Pang

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical studies show that Ver Fed GNN has competitive prediction accuracy with existing privacy preserving GNN frameworks while enhanced privacy protection for users interaction information.
Researcher Affiliation Academia 1Department of Analytics and Operations, National University of Singapore, 119077 Singapore.
Pseudocode Yes Algorithm 1 Federated Vertical Graph Neural Network
Open Source Code Yes Source code: https://github.com/maiph123/Vertical GNN
Open Datasets Yes Dataset: We use two benchmark datasets for recommendation, Movie Lens-1M2 (ML-1M) and Book Crossing3. For Book Crossing we randomly select 6000 users and 3000 items. The items are divided into non-overlapping groups to simulate the vertical federated setting. 2https://grouplens.org/datasets/movielens/1m/ 3http://www2.informatik.uni-freiburg.de/ cziegler/BX/
Dataset Splits Yes Cross-validation is adopted to tune the hyper-parameter, where the training-validation-testing ratio is 60%-20%-20%.
Hardware Specification Yes The experiment is implemented on Ubuntu Linux 20.04 server with 16-core CPU and 64GB RAM, where the programming language is Python.
Software Dependencies No The paper mentions the operating system and programming language: "The experiment is implemented on Ubuntu Linux 20.04 server...programming language is Python." However, it does not specify version numbers for any critical software libraries or frameworks (e.g., PyTorch, TensorFlow, scikit-learn) that would be needed for reproducibility.
Experiment Setup Yes Implementation and Hyper-parameter Setting: Appendix H details the implementation and hyperparameters. Based on the hyper-parameter optimization, we set embedding dimension D to 6, layer size K to 2, learning rate η to 0.05, and neighbor threshold thd to 4 for ML-1M and 8 for Book Crossing. We use sigmoid as the activation function. We consider privacy parameter r from 2 to 50, inverse dimension reduction ratio Nu/q from 1 to 100, and participation rate from 0.2 to 1. The immediate gradients are clipped within [ 0.5, 0.5] so that g1 g2 1 before applying ternary quantization.