Sequential Recommendation System via Self-supervised Learning

Overview

This repo is the code of our survey paper "基于自监督的预训练在推荐系统中的研究综述" accepted by CCIR 2023, which collects several codes and datasets of Self-supervised learning Sequential Recommendation System baselines.

Model	Paper title and link	Code link	Topic	From
ASReP	Augmenting Sequential Recommendation with Pseudo-Prior Items via Reversely Pre-training Transformer	https://github.com/DyGRec/ASReP	Sequential Recommendation	SIGIR2021
SASRec	Self-Attentive Sequential Recommendation	https://github.com/kang205/SASRec	Sequential Recommendation	ICDM2018
DHCN	Self-Supervised Hypergraph Convolutional Networks for Session-based Recommendation	https://github.com/xiaxin1998/DHCN	Session Recommendation	AAAI2021
S3Rec	S3Rec: Self-Supervised Learning for Sequential Recommendation with Mutual Information Maximization	https://github.com/RUCAIBox/CIKM2020-S3Rec	Sequential Recommendation	CIKM2020
MrTransformer	Improving Transformer-based Sequential Recommenders through Preference Editing	https://github.com/mamuyang/MrTransformer	Sequential Recommendation	TOIS2022
BERT4Rec	BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations from Transformer	https://github.com/FeiSun/BERT4Rec	Sequential Recommendation	CIKM2019
CL4SRec	Contrastive Learning for Sequential Recommendation	our reproduction via RecBole and DuoRec	Sequential Recommendation	ICDE2022
SGL	Self-supervised Graph Learning for Recommendation	https://github.com/wujcan/SGL	Session Recommendation	SIGIR2021

For CL4Rec and SGL models, we reproduce them and run experiment with RecBole.

Compared with the original paper, the code is changed, e.g., we have added the code to count the indicators of different length.

Datasets

	Beauty	ML-1M	Yelp
User	22364	6040	22845
Item	12102	3352	16552
Interaction	194687	269721	237004
Total File	4.18M	5.30M	5.19M
Min_len	5	17	5
Max_len	50	50	50
Avg_len	8.7057	44.6557	10.37443
Density	0.07194251%	1.3322134%	0.06267784%
Attributes	2320	18	1158
Min. Attribute / Item	1	1	0
Max. Attribute / Item	9	6	33
Avg. Attribute / Item	3.9391	1.7072	4.9205

length	Beauty	ML-1M	Yelp
[0,20)	21228 | 94.9202%	177 | 2.9305%	20744 | 90.8032%
[20,30)	655 | 2.9289%	684 | 11.3245%	1094 | 4.7888%
[30,40)	231 | 1.0330%	543 | 8.9901%	511 | 2.2368%
[40,50]	250 | 1.1179%	4636 | 76.7550%	496 | 2.1712%
overall	22364 | 100%	6040 | 100%	22845 | 100%

Datasets PreProcessing

We refer to the method in [1,2,3] to process the datasets. If the user interacts with the item, we will convert the interaction with a clear score into implicit positive feedback. After that, we will group the interactive information according to users. We will sort the items for each user according to the timestamp of their interaction with the items. Because this work aims not to investigate the "cold start" issue in the recommendation system, we circularly filter out users with less than 5 interactions and items with less than 5 interactions. In addition, there are users with too much interactive data in the dataset used in this work, so we limit the maximum length of the user interaction sequences to 50. Because the yelp dataset is too large, we adopted a processing method similar to [3], and only the 2019 data of the dataset was intercepted.

Reference:

[1] Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recommendation. In ICDM. IEEE, 197–206.

[2] Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. 2019. BERT4Rec: Sequential recommendation with bidirectional encoder representations from transformer. In CIKM. 1441–1450.

[3] S3-Rec: Self-Supervised Learning for Sequential Recommendation with Mutual Information Maximization. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management (CIKM '20).

Usage

Environments

For each model, the requirement.txt is provided, you can use

pip install -r requirements.txt if you use pip.

conda install --yes --file requirements.txt if you use conda.

Slurm

Besides, we provide the slurm execute script in each baseline folder.

More details about slurm usage can be found on this link: https://slurm.schedmd.com/documentation.html

[Note: please modify "conda activate envname" to your environment]

#!/bin/bash
#SBATCH -e sas_ans_FT.err
#SBATCH -o sas_ans_FT.out
#SBATCH -J sas4recFT # jobname

#SBATCH --partition=debug 
#SBATCH --nodelist=gpuxx
#SBATCH --gres=gpu:1
#SBATCH --cpus-per-task=4
#SBATCH --time=999:00:00


conda activate xxx
python main.py