im0qianqian / LAGCL

Long-tail Augmented Graph Contrastive Learning for Recommendation, ECML/PKDD 2023

LAGCL

Source code for "Long-tail Augmented Graph Contrastive Learning for Recommendation" (ECML/PKDD 2023).

Overview

Experiments

The statistics of three datasets

Dataset	#Users	#Items	#Interactions	Density
Yelp2018	31,668	38,048	1,561,406	0.1296%
Amazon-Book	52,643	91,599	2,984,108	0.0619%
Movielens-25M	155,002	27,133	3,612,474	0.0859%

Overall performance comparison

In Paper

In this repo (using the same train and test data)

	Recall@20 (Paper)	Recall@20 (Recurrence)	NDCG@20 (Paper)	NDCG@20 (Recurrence)
Yelp2018	0.0732	0.07292	0.0604	0.06003
Amazon Book	0.0522	0.05249	0.0415	0.04144
ml-25M	0.3579	0.35938	0.2509	0.25370
LastFM	0.2737	0.30890	0.2721	0.29995

Distribution uniformity analysis

Usage

1. Envs and Dependencies

# create a conda envs
conda create -n lagcl_envs -y python=3.8 ipython
conda activate lagcl_envs

# install pytorch with cuda 11.7
pip install torch==2.0.1 --index-url=https://download.pytorch.org/whl/cu117

# install dgl
pip install dgl==1.1.1 -f https://data.dgl.ai/wheels/cu117/repo.html
pip install dglgo==0.0.2 -f https://data.dgl.ai/wheels-test/repo.html

# install others
pip install numba==0.57.1
pip install wandb==0.15.8

2. Processing for Dataset

Download the ecml_pkdd_2023_lagcl_datasets.zip from here, and unzip it to the datasets/ directory.

The directory tree of datasets/ are below.

.
├── amazon-book
│   ├── test.txt
│   └── train.txt
├── lastfm
│   ├── readme.txt
│   ├── split.py
│   ├── test.txt
│   └── train.txt
├── ml-25m
│   ├── README.md
│   ├── split.py
│   ├── test.txt
│   └── train.txt
└── yelp2018
    ├── process.py
    ├── test.txt
    └── train.txt

3. Training and Evaluation

Run example on Yelp2018 or other datasets with LAGCL Model:

python run.py --dataset_name yelp2018
python run.py --dataset_name amazon-book
python run.py --dataset_name ml-25m
python run.py --dataset_name lastfm

default params

dataset_name	lastfm	yelp2018	amazon-book	ml-25m
batch_size	2048	2048	2048	2048
n_hops	2	2	2	2
user_num	1890	31668	52643	153315
item_num	15387	38048	91599	25121
hidden_size	64	64	64	64
learning_rate	0.001	0.001	0.001	0.001
metric_topk	20	20	20	20
epoch	50	20	15	6
noise_eps	0.1	0.1	0.1	0.2
cl_tau	0.2	0.2	0.2	0.2
cl_rate	0.1	0.5	2	0.2
tail_k_threshold	40	20	30	18
m_h_eps	0.01	0.001	0.001	0.1
kl_eps	10	10	10	0.1
neighbor_norm_type	both	left	left	both
annealing_type	3	0	1	2
agg_function	1	0	0	0
lambda_gp	1	1	1	1
use_relation_rf	FALSE	TRUE	TRUE	TRUE
agg_w_exp_scale	20	20	20	20
l2_reg	0.0001	0.0001	0.0001	0.0001

References

Please cite this repository using the following reference:

@inproceedings{zhao-etal-2023-lagcl,
    title = "Long-tail Augmented Graph Contrastive Learning for Recommendation",
    author = "Zhao, Qian and Wu, Zhengwei and Zhang, Zhiqiang and Zhou, Jun",
    booktitle = "European Conference, ECML PKDD 2023",
    year = "2023",
}