LightGT: A Light Graph Transformer for Multimedia Recommendation

Yinwei Wei¹, Wenqi Liu², Fan Liu¹, Xiang Wang³, Liqiang Nie⁴, Tat-Seng Chua¹

¹National University of Singapore    ²Shandong University   
³University of Science and Technology of China    ⁴Harbin Institute of Technology, Shenzhen   

Updates

• [04/2026] Formatted README.
• [07/2023] Paper accepted at ACM SIGIR 2023, Taipei.
• [07/2023] Code released.

---

Introduction

This is the official PyTorch implementation of LightGT, a Light Graph Transformer for Multimedia Recommendation.

---

Project Structure

.
├── image/                 # Framework and model figures
│   ├── figure1.png
│   └── figure2.png
├── main.py                # Training and evaluation entry point
├── model.py               # LightGT model definition
├── transformer.py         # Transformer module
├── dataloader.py          # Data loading utilities
├── Parser.py              # Argument parser
├── sparsity_group_test.py # Sparsity group evaluation
└── README.md

---

Installation

1. Clone the repository

git clone https://github.com/iLearn-Lab/SIGIR23-LightGT.git
cd SIGIR23-LightGT

2. Install dependencies

The code has been tested under Python 3.8.15. Required packages:

• PyTorch == 1.7.0
• NumPy == 1.23.4

---

Dataset

You can find the full version of recommendation datasets via Kwai, Tiktok, and Movielens. Due to copyright restrictions, we cannot release them directly.

	#Interactions	#Users	#Items	Visual	Acoustic	Textual
Movielens	1,239,508	55,485	5,986	2,048	128	100
Tiktok	726,065	36,656	76,085	128	128	128
Kwai	1,664,305	22,611	329,510	2,048	-	100

MMGCN provides corresponding toy datasets that can be used for research.

Data format:

• train.npy — Train file. Each line is a user with positive interactions: (userID, itemID)
• val.npy — Validation file. Each line is a user with positive interactions: (userID, itemID)
• test.npy — Test file. Each line is a user with positive interactions: (userID, itemID)

---

Usage

Training & Evaluation

• Movielens dataset

  python main.py --l_r=1e-2 --weight_decay=1e-2 --src_len=50 --score_weight=0.05 --nhead=1 --transformer_layers=4 --batch_size=2048 --lightgcn_layers=4 --dataset=movielens

• Tiktok dataset

  python main.py --l_r=1e-2 --weight_decay=1e-2 --src_len=50 --score_weight=0.05 --nhead=1 --transformer_layers=4 --batch_size=2048 --lightgcn_layers=4 --dataset=tiktok

• Kwai dataset

  python main.py --l_r=1e-2 --weight_decay=1e-2 --src_len=50 --score_weight=0.05 --nhead=1 --transformer_layers=4 --batch_size=2048 --lightgcn_layers=4 --dataset=kwai

---

Citation

If you find this work useful for your research, please kindly cite our paper:

@inproceedings{wei2023lightgt,
  title      = {Lightgt: A light graph transformer for multimedia recommendation},
  author     = {Wei, Yinwei and
                Liu, Wenqi and
                Liu, Fan and
                Wang, Xiang and
                Nie, Liqiang and
                Chua, Tat-Seng},
  booktitle  = {Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval},
  pages      = {1508--1517},
  year       = {2023}
}

---

Acknowledgement

This work is developed based on MMGCN and LightGCN. We thank the authors for their open-source contributions.