Skip to content

README.md

Latest commit

 

History

History
456 lines (329 loc) · 10.8 KB

README.md

File metadata and controls

456 lines (329 loc) · 10.8 KB

R2NS

中文 | English

WWW 2026 paper: "R2NS: Recall and Re-ranking of Negative Samples for Sequential Recommendation"

中文说明

项目简介

本仓库是 WWW 2026 论文 R2NS: Recall and Re-ranking of Negative Samples for Sequential Recommendation 的官方代码仓库。

仓库主要包含三部分：

  • Ours/：R2NS 主方法实现
  • Baselines/：多种负采样基线方法
  • WWW_2026_Camera_Ready.pdf：论文 camera-ready 版本

快速跳转

中文仓库结构

R2NS/
├── Ours/
│   ├── data/
│   ├── run_finetune.bash
│   ├── run_finetune_full.py
│   ├── datasets.py
│   ├── models.py
│   ├── modules.py
│   ├── trainers.py
│   └── utils.py
├── Baselines/
│   ├── Neg_samples_DNS+/
│   ├── Neg_samples_gnno/
│   ├── Neg_samples_posmix/
│   ├── Neg_samples_srns/
│   └── Neg_samples_two_pass/
├── WWW_2026_Camera_Ready.pdf
└── README.md

中文环境依赖

仓库目前没有提供固定版本的 requirements.txt，需要手动安装依赖。

建议环境：

  • Python 3.9+
  • 支持 CUDA 的 PyTorch
  • NVIDIA GPU

代码中实际使用到的核心依赖包括：

  • torch
  • numpy
  • scipy
  • pandas
  • scikit-learn
  • tqdm
  • openpyxl
  • texttable
  • transformers
  • recbole
  • mamba-ssm

一个常见的环境配置方式如下：

conda create -n r2ns python=3.10 -y
conda activate r2ns

# 请根据你的 CUDA 版本先安装 PyTorch
# 例如：
# pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

pip install numpy scipy pandas scikit-learn tqdm openpyxl texttable transformers recbole mamba-ssm

中文数据格式

代码默认每个数据集包含三个切分文件：

  • DATASET_train.txt
  • DATASET_val.txt
  • DATASET_test.txt

每一行表示一个用户的交互序列，物品 ID 之间用空格分隔，例如：

1802 1314
1802 1314 1313

当前仓库中只附带了 Beauty 的样例数据：

  • Ours/data/Beauty_train.txt
  • Ours/data/Beauty_val.txt
  • Ours/data/Beauty_test.txt

如果你要接入自己的数据集，只需要把三个文件放进 data/ 目录，并保持命名方式不变。

中文主方法运行

单次运行

请从 Ours/ 目录启动：

cd Ours

python run_finetune_full.py \
  --data_name Beauty \
  --backbone SASRec \
  --stage2_epoch 30 \
  --neg_sampler Uniform \
  --CL_type Gentle \
  --loss_type BCE \
  --M 5 \
  --N 100 \
  --K1 0.05

代码和脚本中常见的 backbone 包括：

  • SASRec
  • Mamba4Rec
  • GRU4Rec
  • LightSANs
  • Linrec
  • FMLPRecModel

批量运行

cd Ours
bash run_finetune.bash

运行批处理脚本前，建议先检查并修改：

  • models
  • datasets
  • K1
  • stage2_epoch
  • AVAILABLE_GPUS
  • JOBS_PER_GPU

该脚本会并行启动多个训练任务，并将日志写入脚本中配置的输出目录。

中文基线运行

仓库中包含以下基线目录：

  • Baselines/Neg_samples_DNS+
  • Baselines/Neg_samples_gnno
  • Baselines/Neg_samples_posmix
  • Baselines/Neg_samples_srns
  • Baselines/Neg_samples_two_pass

每个基线目录都带有各自的 run_finetune_full.py 和 run_finetune.bash。

重要：基线的数据路径

基线脚本默认从 ./data/ 读取数据，但当前仓库里实际只有 Ours/data/。

因此在运行基线前，你需要做下面两件事中的一种：

  1. 在对应基线目录下创建一个名为 data 的软链接。
  2. 手动运行 run_finetune_full.py，并显式传入 --data_dir ../../Ours/data/。

示例：

ln -s ../../Ours/data Baselines/Neg_samples_srns/data

cd Baselines/Neg_samples_srns
bash run_finetune.bash

或者：

cd Baselines/Neg_samples_srns

python run_finetune_full.py \
  --data_dir ../../Ours/data/ \
  --data_name Beauty \
  --backbone SASRec \
  --neg_sampler DNS \
  --CL_type Radical \
  --start_epoch 30 \
  --N 100

中文输出结果

根据不同脚本和目录，程序通常会生成：

  • 实验日志目录，例如 res/、res2/ 或其他自定义目录
  • 保存在 output/ 下的模型 checkpoint
  • 类似 CL1.xlsx 的 Excel 汇总文件

中文复现注意事项

这个仓库已经足够用于阅读和继续实验，但如果你准备大规模复现，建议先检查以下几点：

  • Ours/run_finetune_full.py 在写结果时使用了 args.start2_epoch，但参数解析里实际定义的是 stage2_epoch。
  • Ours/utils.py 里写 Excel 时使用了 args.K，而主方法脚本中定义的是 K1 和 K2。
  • 多个批量脚本里带有手工覆盖过的数据集列表、模型列表、GPU ID 和输出目录。
  • 当前仓库快照里只包含 Ours/data/Beauty_* 这组示例数据。

如果你要严格复现论文最终表格，建议先检查一遍脚本参数和结果落盘逻辑，再启动长时间训练任务。

中文引用说明

如果这个仓库对你的研究有帮助，请引用对应的 WWW 2026 论文。

English

Overview

This repository is the official implementation of the WWW 2026 paper "R2NS: Recall and Re-ranking of Negative Samples for Sequential Recommendation".

The repository mainly contains:

  • Ours/: the implementation of R2NS
  • Baselines/: several negative sampling baselines
  • WWW_2026_Camera_Ready.pdf: the camera-ready paper

Quick Links

Repository Structure

R2NS/
├── Ours/
│   ├── data/
│   ├── run_finetune.bash
│   ├── run_finetune_full.py
│   ├── datasets.py
│   ├── models.py
│   ├── modules.py
│   ├── trainers.py
│   └── utils.py
├── Baselines/
│   ├── Neg_samples_DNS+/
│   ├── Neg_samples_gnno/
│   ├── Neg_samples_posmix/
│   ├── Neg_samples_srns/
│   └── Neg_samples_two_pass/
├── WWW_2026_Camera_Ready.pdf
└── README.md

Environment

This repository does not currently provide a pinned requirements.txt, so dependencies need to be installed manually.

Recommended environment:

  • Python 3.9+
  • PyTorch with CUDA support
  • NVIDIA GPU

Core dependencies used by the code include:

  • torch
  • numpy
  • scipy
  • pandas
  • scikit-learn
  • tqdm
  • openpyxl
  • texttable
  • transformers
  • recbole
  • mamba-ssm

A typical setup is:

conda create -n r2ns python=3.10 -y
conda activate r2ns

# Install PyTorch first according to your CUDA version
# Example:
# pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

pip install numpy scipy pandas scikit-learn tqdm openpyxl texttable transformers recbole mamba-ssm

Data Format

Each dataset is expected to contain three split files:

  • DATASET_train.txt
  • DATASET_val.txt
  • DATASET_test.txt

Each line is a user interaction sequence represented by space-separated item IDs, for example:

1802 1314
1802 1314 1313

The current repository snapshot only includes example files for Beauty:

  • Ours/data/Beauty_train.txt
  • Ours/data/Beauty_val.txt
  • Ours/data/Beauty_test.txt

To use your own dataset, place the three files under a data/ directory and keep the naming convention unchanged.

Run R2NS

Single run

Run from the Ours/ directory:

cd Ours

python run_finetune_full.py \
  --data_name Beauty \
  --backbone SASRec \
  --stage2_epoch 30 \
  --neg_sampler Uniform \
  --CL_type Gentle \
  --loss_type BCE \
  --M 5 \
  --N 100 \
  --K1 0.05

Common backbone choices in the codebase include:

  • SASRec
  • Mamba4Rec
  • GRU4Rec
  • LightSANs
  • Linrec
  • FMLPRecModel

Batch run

cd Ours
bash run_finetune.bash

Before launching the batch script, you should review:

  • models
  • datasets
  • K1
  • stage2_epoch
  • AVAILABLE_GPUS
  • JOBS_PER_GPU

The script launches multiple jobs in parallel and writes logs to its configured output directory.

Run Baselines

The repository includes the following baseline directories:

  • Baselines/Neg_samples_DNS+
  • Baselines/Neg_samples_gnno
  • Baselines/Neg_samples_posmix
  • Baselines/Neg_samples_srns
  • Baselines/Neg_samples_two_pass

Each baseline directory has its own run_finetune_full.py and run_finetune.bash.

Important: data path for baselines

Baseline scripts read data from ./data/ by default, while this repository currently only includes Ours/data/.

So before running a baseline, do one of the following:

  1. Create a symbolic link named data inside the target baseline directory.
  2. Run run_finetune_full.py manually and pass --data_dir ../../Ours/data/.

Example:

ln -s ../../Ours/data Baselines/Neg_samples_srns/data

cd Baselines/Neg_samples_srns
bash run_finetune.bash

Or:

cd Baselines/Neg_samples_srns

python run_finetune_full.py \
  --data_dir ../../Ours/data/ \
  --data_name Beauty \
  --backbone SASRec \
  --neg_sampler DNS \
  --CL_type Radical \
  --start_epoch 30 \
  --N 100

Outputs

Depending on the script and directory, the code may generate:

  • experiment log directories such as res/, res2/, or other custom folders
  • model checkpoints under output/
  • Excel summary files such as CL1.xlsx

Reproducibility Notes

This repository is usable for reading and continuing experiments, but a few details should be checked before large-scale reproduction:

  • Ours/run_finetune_full.py writes results using args.start2_epoch, while the parser actually defines stage2_epoch.
  • Ours/utils.py expects args.K when writing Excel results, while the main R2NS script defines K1 and K2.
  • Several batch scripts contain manually overwritten dataset lists, model lists, GPU IDs, and output directories.
  • The current repository snapshot only includes Ours/data/Beauty_*.

If you plan to reproduce the final tables strictly, review the script arguments and result-writing logic before launching long runs.

Citation

If you find this repository useful, please cite the corresponding WWW 2026 paper.