Skip to content

zerolovesea/NextRec

PyPI Downloads Python PyTorch License Version Ask DeepWiki

中文文档 | English Version

统一、高效、可扩展的推荐系统框架

目录

简介

NextRec是一个基于PyTorch的现代推荐系统框架,旨在为研究工程团队提供快速的建模、训练与评估流。框架内置丰富的模型库、数据处理工具和工程化训练组件。此外提供了易上手的接口,命令行工具及教程,推荐算法学习者能以最快速度了解模型架构,复现学术论文并进行训练和部署。

Why NextRec

  • 多场景推荐能力:覆盖排序(CTR/CVR)、召回、多任务学习、生成式召回等推荐/营销模型,持续跟进业界进展。
  • 统一的特征工程与数据流水线:NextRec框架提供了统一的特征定义、可持久化的数据处理、批处理优化,符合工业大数据Spark/Hive场景下,基于离线特征的模型训练推理流程。
  • 友好的工程体验:支持多种格式数据(csv/parquet/pathlike)的流式处理/分布式训练/推理与可视化指标监控,方便业务算法工程师和推荐算法学习者快速复现实验。
  • 灵活的命令行工具:支持通过命令行和配置文件,一键启动训练和推理进程,方便快速实验迭代和敏捷部署。
  • 高效训练与评估:内置多种优化器、学习率调度、早停、模型检查点与详细的日志管理,开箱即用。

NextRec近期进展

  • 28/12/2025 在v0.4.21中加入了对SwanLab和Wandb的支持,通过model的fit方法进行配置:use_swanlab=True, swanlab_kwargs={"project": "NextRec","name":"tutorial_movielens_deepfm"},
  • 21/12/2025 在v0.4.16中加入了对GradNorm的支持,通过compile的loss_weight='grad_norm'进行配置
  • 12/12/2025 在v0.4.9中加入了RQ-VAE模块。配套的数据集代码已经同步在仓库中
  • 07/12/2025 发布了NextRec CLI命令行工具,它允许用户根据配置文件进行一键训练和推理,我们提供了相关的教程教学代码
  • 03/12/2025 NextRec获得了100颗🌟!感谢大家的支持
  • 06/12/2025 在v0.4.1中支持了单机多卡的分布式DDP训练,并且提供了配套的代码
  • 11/11/2025 NextRec v0.1.0发布,我们提供了10余种Ranking模型,4种多任务模型和4种召回模型,以及统一的训练/日志/指标管理系统

架构

NextRec采用模块化工程设计,核心组件包括:统一特征驱动的BaseModel架构;独立Layer模块;支持训练推理的统一的DataLoader;命令行工具NextCLI等。

NextRec架构

安装

开发者可以通过pip install nextrec快速安装NextRec的最新版本,环境要求为Python 3.10+(对于需要使用CUDA加速的开发者,建议安装对应版本的pytorch)。如果需要执行示例代码,则需要先拉取仓库:

git clone https://github.com/zerolovesea/NextRec.git
cd NextRec/
pip install nextrec # or pip install -e .

示例代码

我们在tutorials/ 目录提供了多个示例,覆盖排序、召回、多任务、数据处理等场景:

如果想了解更多NextRec框架的细节,我们还提供了Jupyter notebook来帮助你了解:

5分钟快速上手

我们提供了详细的上手指南和配套数据集,帮助您熟悉NextRec框架的不同功能。我们在datasets/路径下提供了一个来自电商场景的测试数据集,数据示例如下:

user_id item_id dense_0 dense_1 dense_2 dense_3 dense_4 dense_5 dense_6 dense_7 sparse_0 sparse_1 sparse_2 sparse_3 sparse_4 sparse_5 sparse_6 sparse_7 sparse_8 sparse_9 sequence_0 sequence_1 label
1 7817 0.14704075 0.31020382 0.77780896 0.944897 0.62315375 0.57124174 0.77009535 0.3211029 315 260 379 146 168 161 138 88 5 312 [170,175,97,338,105,353,272,546,175,545,463,128,0,0,0] [368,414,820,405,548,63,327,0,0,0,0,0,0,0,0] 0
1 3579 0.77811223 0.80359334 0.5185201 0.91091245 0.043562356 0.82142705 0.8803686 0.33748195 149 229 442 6 167 252 25 402 7 168 [179,48,61,551,284,165,344,151,0,0,0,0,0,0,0] [814,0,0,0,0,0,0,0,0,0,0,0,0,0,0] 1

接下来我们将用一个简短的示例,展示如何使用NextRec训练一个DIN(Deep Interest Network)模型。您也可以直接执行python tutorials/example_ranking_din.py来执行训练推理代码。

开始训练以后,你可以在nextrec_logs/din_tutorial路径下查看详细的训练日志。

import pandas as pd
from nextrec.models.ranking.din import DIN
from nextrec.basic.features import DenseFeature, SparseFeature, SequenceFeature

df = pd.read_csv('dataset/ranking_task.csv')

for col in df.columns and 'sequence' in col: # csv默认将列表读取成文本,我们需要将其转化为对象
    df[col] = df[col].apply(lambda x: eval(x) if isinstance(x, str) else x)

# 我们需要将不同特征进行定义
dense_features = [DenseFeature(name=f'dense_{i}', input_dim=1) for i in range(8)]

sparse_features = [SparseFeature(name='user_id', embedding_name='user_emb', vocab_size=int(df['user_id'].max() + 1), embedding_dim=32), SparseFeature(name='item_id', embedding_name='item_emb', vocab_size=int(df['item_id'].max() + 1), embedding_dim=32),]

sparse_features.extend([SparseFeature(name=f'sparse_{i}', embedding_name=f'sparse_{i}_emb', vocab_size=int(df[f'sparse_{i}'].max() + 1), embedding_dim=32) for i in range(10)])

sequence_features = [
    SequenceFeature(name='sequence_0', vocab_size=int(df['sequence_0'].apply(lambda x: max(x)).max() + 1), embedding_dim=32, padding_idx=0, embedding_name='item_emb'),
    SequenceFeature(name='sequence_1', vocab_size=int(df['sequence_1'].apply(lambda x: max(x)).max() + 1), embedding_dim=16, padding_idx=0, embedding_name='sparse_0_emb'),]

mlp_params = {
    "dims": [256, 128, 64],
    "activation": "relu",
    "dropout": 0.3,
}

model = DIN(
    dense_features=dense_features,
    sparse_features=sparse_features,
    sequence_features=sequence_features,
    mlp_params=mlp_params,
    attention_hidden_units=[80, 40],
    attention_activation='sigmoid',
    attention_use_softmax=True,
    target=['label'],                                     # 目标变量
    device='mps',                                         
    embedding_l1_reg=1e-6,
    embedding_l2_reg=1e-5,
    dense_l1_reg=1e-5,
    dense_l2_reg=1e-4,
    session_id="din_tutorial",                            # 实验id,用于存放训练日志
)

# 编译模型,设置优化器和损失函数
model.compile(
            optimizer = "adam",
            optimizer_params = {"lr": 1e-3, "weight_decay": 1e-5},
            loss = "focal",
            loss_params={"gamma": 2.0, "alpha": 0.25},
        )

model.fit(
    train_data=df,
    metrics=['auc', 'gauc', 'logloss'],  # 添加需要查看的指标
    epochs=3,
    batch_size=512,
    shuffle=True,
    user_id_column='user_id'             # 用于计算GAUC的id列 
)

# 训练完成后进行指标评估
metrics = model.evaluate(
    df,
    metrics=['auc', 'gauc', 'logloss'],
    batch_size=512,
    user_id_column='user_id'
)

命令行工具

NextRec 提供了强大的命令行界面,支持通过 YAML 配置文件进行模型训练和预测。详细的 CLI 文档请参见:

# 训练模型
nextrec --mode=train --train_config=path/to/train_config.yaml

# 运行预测
nextrec --mode=predict --predict_config=path/to/predict_config.yaml

预测结果固定保存到 {checkpoint_path}/predictions/{name}.{save_data_format}

截止当前版本0.4.21,NextRec CLI支持单机训练,分布式训练相关功能尚在开发中。

兼容平台

当前最新版本为0.4.21,所有模型和测试代码均已在以下平台通过验证,如果开发者在使用中遇到兼容问题,请在issue区提出错误报告及系统版本:

平台 配置
MacOS latest MacBook Pro M4 Pro 24G内存
Ubuntu latest AutoDL 4070D 双卡
Ubuntu 24.04 NVIDIA TITAN V 5卡
CentOS 7 Intel Xeon 5138Y 96核 377G内存

支持模型

排序模型

模型 论文 年份 状态
FM Factorization Machines ICDM 2010 已支持
LR Logistic Regression - 已支持
AFM Attentional Factorization Machines: Learning the Weight of Feature Interactions via Attention Networks IJCAI 2017 已支持
FFM Field-aware Factorization Machines RecSys 2016 已支持
DeepFM DeepFM: A Factorization-Machine based Neural Network for CTR Prediction IJCAI 2017 已支持
Wide&Deep Wide & Deep Learning for Recommender Systems DLRS 2016 已支持
xDeepFM xDeepFM: Combining Explicit and Implicit Feature Interactions KDD 2018 已支持
FiBiNET FiBiNET: Combining Feature Importance and Bilinear Feature Interaction for CTR Prediction RecSys 2019 已支持
PNN Product-based Neural Networks for User Response Prediction ICDM 2016 已支持
AutoInt AutoInt: Automatic Feature Interaction Learning CIKM 2019 已支持
DCN Deep & Cross Network for Ad Click Predictions ADKDD 2017 已支持
DCN v2 DCN V2: Improved Deep & Cross Network and Practical Lessons for Web-scale Learning to Rank Systems KDD 2021 已支持
DIN Deep Interest Network for Click-Through Rate Prediction KDD 2018 已支持
DIEN Deep Interest Evolution Network for Click-Through Rate Prediction AAAI 2019 已支持
MaskNet MaskNet: Introducing Feature-wise Gating Blocks for High-dimensional Sparse Recommendation Data 2020 已支持
EulerNet EulerNet: Efficient and Effective Feature Interaction Modeling with Euler's Formula SIGIR 2021 已支持

召回模型

模型 论文 年份 状态
DSSM Learning Deep Structured Semantic Models CIKM 2013 已支持
DSSM v2 DSSM with pairwise BPR-style optimization - 已支持
YouTube DNN Deep Neural Networks for YouTube Recommendations RecSys 2016 已支持
MIND Multi-Interest Network with Dynamic Routing CIKM 2019 已支持
SDM Sequential Deep Matching Model - 已支持

序列推荐模型

模型 论文 年份 状态
SASRec Self-Attentive Sequential Recommendation KDD 2018 开发中
HSTU Actions speak louder than words: Trillion-parameter sequential transducers for generative recommendations arXiv 2024 已支持

多任务模型

模型 论文 年份 状态
MMOE Modeling Task Relationships in Multi-task Learning KDD 2018 已支持
PLE Progressive Layered Extraction RecSys 2020 已支持
ESMM Entire Space Multi-Task Model SIGIR 2018 已支持
ShareBottom Multitask Learning - 已支持
POSO POSO: Personalized Cold-start Modules for Large-scale Recommender Systems 2021 已支持

生成式模型

模型 论文 年份 状态
TIGER Recommender Systems with Generative Retrieval NeurIPS 2023 开发中

表征模型

模型 论文 年份 状态
RQ-VAE RQ-VAE: RQVAE for Generative Retrieval - 已支持
BPR Bayesian Personalized Ranking UAI 2009 开发中
MF Matrix Factorization Techniques for Recommender Systems - 开发中
AutoRec AutoRec: Autoencoders Meet Collaborative Filtering WWW 2015 开发中
LightGCN LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation SIGIR 2020 开发中
S3Rec S3-Rec: Self-Supervised Learning for Sequential Recommendation CIKM 2020 开发中
CL4SRec CL4SRec: Contrastive Learning for Sequential Recommendation 2021 开发中

贡献指南

我们欢迎任何形式的贡献!

如何贡献

  1. Fork 本仓库
  2. 创建特性分支 (git checkout -b feature/AmazingFeature)
  3. 提交更改 (git commit -m 'Add some AmazingFeature')
  4. 推送到分支 (git push origin feature/AmazingFeature)
  5. 创建 Pull Request

在提交 PR 之前,请运行 python test/run_tests.pypython scripts/format_code.py 确保所有测试通过并统一代码风格。

代码规范

  • 遵循 PEP 8 Python 代码风格
  • 为新增功能补充单元测试
  • 同步更新相关文档

报告错误

Issues 页面提交问题时,请包含:

  • 错误描述
  • 重现步骤
  • 期望行为
  • 实际行为
  • 环境信息(Python 版本、PyTorch 版本等)

许可证

本项目采用 Apache 2.0 许可证

联系方式

致谢

NextRec 的开发受到以下优秀项目的启发:

  • torch-rechub - 灵活且易于扩展的推荐系统框架
  • FuxiCTR - 可配置、可调优、可复现的 CTR 预测库
  • RecBole - 统一、全面、高效的推荐库

感谢开源社区的所有贡献者!

引用

如果您在研究或工作中使用了本框架,欢迎引用本项目:

@misc{nextrec,
    title = {NextRec},
    author = {Yang Zhou},
    year = {2025},
    publisher = {GitHub},
    journal = {GitHub repository},
    howpublished = {\url{https://github.com/zerolovesea/NextRec}},
    note = {A unified, efficient, and extensible PyTorch-based recommendation library}
}

About

A unified, efficient, and extensible PyTorch-based recommendation library

Topics

Resources

License

Code of conduct

Contributing

Stars

Watchers

Forks

Packages

No packages published