MCLRec

This is our Pytorch implementation for the paper: "Meta-optimized Contrastive Learning for Sequential Recommendation".

Environment Requirement

Pytorch>=1.7.0
Python>=3.7

Model Overview

Usage

Please run the following command to install all the requirements:

pip install -r requirements.txt

Datasets Prepare

Please use the data_process.py under dataset/ to get the input dataset by running the following command :

python data_process.py

Evaluate Model

We provide the trained models on Amazon_Beauty, Amazon_Sports_and_Outdoors, and Yelp datasets in ./log/Checkpoint/<Data_name>folder. You can directly evaluate the trained models on test set by running:

python run_seq.py --dataset=<Data_name> --do_eval

On Amazon_Beauty:

python run_seq.py --dataset=Amazon_Beauty --do_eval

 INFO  test result: {'recall@5': 0.0581, 'recall@10': 0.0871, 'recall@20': 0.1243, 'recall@50': 0.1852, 'mrr@5': 0.0278, 'mrr@10': 0.0316, 'mrr@20': 0.0341, 'mrr@50': 0.036, 'ndcg@5': 0.0352, 'ndcg@10': 0.0446, 'ndcg@20': 0.0539, 'ndcg@50': 0.066, 'precision@5': 0.0116, 'precision@10': 0.0087, 'precision@20': 0.0062, 'precision@50': 0.0037}

On Amazon_Sports_and_Outdoors:

python run_seq.py --dataset=Amazon_Sports_and_Outdoors --do_eval

INFO  test result: {'recall@5': 0.0328, 'recall@10': 0.0501, 'recall@20': 0.0734, 'recall@50': 0.1215, 'mrr@5': 0.0163, 'mrr@10': 0.0186, 'mrr@20': 0.0202, 'mrr@50': 0.0218, 'ndcg@5': 0.0204, 'ndcg@10': 0.026, 'ndcg@20': 0.0319, 'ndcg@50': 0.0414, 'precision@5': 0.0066, 'precision@10': 0.005, 'precision@20': 0.0037, 'precision@50': 0.0024}

On Yelp:

python run_seq.py --dataset=Yelp --do_eval

INFO  test result: {'recall@5': 0.0454, 'recall@10': 0.0647, 'recall@20': 0.0941, 'recall@50': 0.1557, 'mrr@5': 0.0292, 'mrr@10': 0.0317, 'mrr@20': 0.0337, 'mrr@50': 0.0356, 'ndcg@5': 0.0332, 'ndcg@10': 0.0394, 'ndcg@20': 0.0467, 'ndcg@50': 0.0589, 'precision@5': 0.0091, 'precision@10': 0.0065, 'precision@20': 0.0047, 'precision@50': 0.0031}

Train Model

Please train the model using the Python script run_seq.py.

You can run the following command to train the model on Yelp datasets:

python run_seq.py --dataset=Yelp --epochs=100 --use_rl=1 --joint=0 train_batch_size=256 --lmd=0.03 --beta=0.1 --sim='dot'

or You can cd scripts and run the following command to train the model on different dataset:

bash scrips/train_{dataset name}.sh

bash scripts/train_beauty.sh
bash scripts/train_ml-1m.sh
bash scripts/train_sports.sh
bash scripts/train_toys.sh
bash scripts/train_yelp.sh

Overall Performances

N represents Normalized Discounted Cumulative Gain(NDCG) and H represents Hit Ratio (HR).

Dataset	Metrc	BPR	GRU4Rec	Caser	SASRec	BERT4Rec	S3Rec	CL4SRec	CoSeRec	LMA4Rec	ICLRec	DuoRec	SRMA	MCLRec	Improv.
Sports	H@5	0.0123	0.0162	0.0154	0.0214	0.0217	0.0121	0.0231	0.0290	0.0297	0.0290	0.0312	0.0299	0.0328	5.13%
Sports	H@20	0.0369	0.0421	0.0399	0.0500	0.0604	0.0344	0.0557	0.0636	0.0634	0.0646	0.0696	0.0649	0.0734	5.46%
Sports	N@5	0.0076	0.0103	0.0114	0.0144	0.0143	0.0084	0.0146	0.0196	0.0197	0.0191	0.0192	0.0199	0.0204	2.51%
Sports	N@20	0.0144	0.0186	0.0178	0.0224	0.0251	0.0146	0.0238	0.0293	0.0293	0.0291	0.0302	0.0297	0.0319	5.63%
Beauty	H@5	0.0178	0.0180	0.0251	0.0377	0.0360	0.0189	0.0401	0.0504	0.0511	0.0500	0.0559	0.0503	0.0581	3.94%
Beauty	H@20	0.0474	0.0427	0.0643	0.0894	0.0984	0.0487	0.0974	0.1034	0.1047	0.1058	0.1193	0.1025	0.1243	4.19%
Beauty	N@5	0.0109	0.0116	0.0145	0.0241	0.0216	0.0115	0.0268	0.0339	0.0342	0.0326	0.0340	0.0318	0.0352	2.92%
Beauty	N@20	0.0192	0.0186	0.0298	0.0386	0.0391	0.0198	0.0428	0.0487	0.0493	0.0483	0.0518	0.0474	0.0539	4.05%
Yelp	H@5	0.0127	0.0152	0.0142	0.0160	0.0196	0.0101	0.0227	0.0241	0.0233	0.0239	0.0429	0.0243	0.0454	5.83%
Yelp	H@20	0.0346	0.0371	0.0406	0.0443	0.0564	0.0314	0.0623	0.0649	0.0636	0.0659	0.0868	0.0646	0.0941	8.41%
Yelp	N@5	0.0082	0.0091	0.0080	0.0101	0.0121	0.0068	0.0143	0.0151	0.0147	0.0152	0.0324	0.0154	0.0332	2.47%
Yelp	N@20	0.0143	0.0145	0.0156	0.0179	0.0223	0.0127	0.0254	0.0263	0.0258	0.0270	0.0447	0.0266	0.0467	4.47%

Future Work

Add MCL on Graph Recommendation
Building Augmenters Using Different Structures

Acknowledgment

Transformer and training pipeline are implemented based on Recbole. Thanks them for providing efficient implementation.

Citation

@inproceedings{MCLRec,
  author       = {Xiuyuan Qin and
                  Huanhuan Yuan and
                  Pengpeng Zhao and
                  Junhua Fang and
                  Fuzhen Zhuang and
                  Guanfeng Liu and
                  Yanchi Liu and
                  Victor S. Sheng},
  title        = {Meta-optimized Contrastive Learning for Sequential Recommendation},
  booktitle    = {SIGIR},
  pages        = {89--98},
  year         = {2023},
}

Please kindly cite our paper if this paper and the code are helpful.

QinHsiu / MCLRec