[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"similar-LAMDA-Tabular--TALENT":3,"tool-LAMDA-Tabular--TALENT":64},[4,17,27,35,43,56],{"id":5,"name":6,"github_repo":7,"description_zh":8,"stars":9,"difficulty_score":10,"last_commit_at":11,"category_tags":12,"status":16},3808,"stable-diffusion-webui","AUTOMATIC1111\u002Fstable-diffusion-webui","stable-diffusion-webui 是一个基于 Gradio 构建的网页版操作界面，旨在让用户能够轻松地在本地运行和使用强大的 Stable Diffusion 图像生成模型。它解决了原始模型依赖命令行、操作门槛高且功能分散的痛点，将复杂的 AI 绘图流程整合进一个直观易用的图形化平台。\n\n无论是希望快速上手的普通创作者、需要精细控制画面细节的设计师，还是想要深入探索模型潜力的开发者与研究人员，都能从中获益。其核心亮点在于极高的功能丰富度：不仅支持文生图、图生图、局部重绘（Inpainting）和外绘（Outpainting）等基础模式，还独创了注意力机制调整、提示词矩阵、负向提示词以及“高清修复”等高级功能。此外，它内置了 GFPGAN 和 CodeFormer 等人脸修复工具，支持多种神经网络放大算法，并允许用户通过插件系统无限扩展能力。即使是显存有限的设备，stable-diffusion-webui 也提供了相应的优化选项，让高质量的 AI 艺术创作变得触手可及。",162132,3,"2026-04-05T11:01:52",[13,14,15],"开发框架","图像","Agent","ready",{"id":18,"name":19,"github_repo":20,"description_zh":21,"stars":22,"difficulty_score":23,"last_commit_at":24,"category_tags":25,"status":16},1381,"everything-claude-code","affaan-m\u002Feverything-claude-code","everything-claude-code 是一套专为 AI 编程助手（如 Claude Code、Codex、Cursor 等）打造的高性能优化系统。它不仅仅是一组配置文件，而是一个经过长期实战打磨的完整框架，旨在解决 AI 代理在实际开发中面临的效率低下、记忆丢失、安全隐患及缺乏持续学习能力等核心痛点。\n\n通过引入技能模块化、直觉增强、记忆持久化机制以及内置的安全扫描功能，everything-claude-code 能显著提升 AI 在复杂任务中的表现，帮助开发者构建更稳定、更智能的生产级 AI 代理。其独特的“研究优先”开发理念和针对 Token 消耗的优化策略，使得模型响应更快、成本更低，同时有效防御潜在的攻击向量。\n\n这套工具特别适合软件开发者、AI 研究人员以及希望深度定制 AI 工作流的技术团队使用。无论您是在构建大型代码库，还是需要 AI 协助进行安全审计与自动化测试，everything-claude-code 都能提供强大的底层支持。作为一个曾荣获 Anthropic 黑客大奖的开源项目，它融合了多语言支持与丰富的实战钩子（hooks），让 AI 真正成长为懂上",138956,2,"2026-04-05T11:33:21",[13,15,26],"语言模型",{"id":28,"name":29,"github_repo":30,"description_zh":31,"stars":32,"difficulty_score":23,"last_commit_at":33,"category_tags":34,"status":16},2271,"ComfyUI","Comfy-Org\u002FComfyUI","ComfyUI 是一款功能强大且高度模块化的视觉 AI 引擎，专为设计和执行复杂的 Stable Diffusion 图像生成流程而打造。它摒弃了传统的代码编写模式，采用直观的节点式流程图界面，让用户通过连接不同的功能模块即可构建个性化的生成管线。\n\n这一设计巧妙解决了高级 AI 绘图工作流配置复杂、灵活性不足的痛点。用户无需具备编程背景，也能自由组合模型、调整参数并实时预览效果，轻松实现从基础文生图到多步骤高清修复等各类复杂任务。ComfyUI 拥有极佳的兼容性，不仅支持 Windows、macOS 和 Linux 全平台，还广泛适配 NVIDIA、AMD、Intel 及苹果 Silicon 等多种硬件架构，并率先支持 SDXL、Flux、SD3 等前沿模型。\n\n无论是希望深入探索算法潜力的研究人员和开发者，还是追求极致创作自由度的设计师与资深 AI 绘画爱好者，ComfyUI 都能提供强大的支持。其独特的模块化架构允许社区不断扩展新功能，使其成为当前最灵活、生态最丰富的开源扩散模型工具之一，帮助用户将创意高效转化为现实。",107662,"2026-04-03T11:11:01",[13,14,15],{"id":36,"name":37,"github_repo":38,"description_zh":39,"stars":40,"difficulty_score":23,"last_commit_at":41,"category_tags":42,"status":16},3704,"NextChat","ChatGPTNextWeb\u002FNextChat","NextChat 是一款轻量且极速的 AI 助手，旨在为用户提供流畅、跨平台的大模型交互体验。它完美解决了用户在多设备间切换时难以保持对话连续性，以及面对众多 AI 模型不知如何统一管理的痛点。无论是日常办公、学习辅助还是创意激发，NextChat 都能让用户随时随地通过网页、iOS、Android、Windows、MacOS 或 Linux 端无缝接入智能服务。\n\n这款工具非常适合普通用户、学生、职场人士以及需要私有化部署的企业团队使用。对于开发者而言，它也提供了便捷的自托管方案，支持一键部署到 Vercel 或 Zeabur 等平台。\n\nNextChat 的核心亮点在于其广泛的模型兼容性，原生支持 Claude、DeepSeek、GPT-4 及 Gemini Pro 等主流大模型，让用户在一个界面即可自由切换不同 AI 能力。此外，它还率先支持 MCP（Model Context Protocol）协议，增强了上下文处理能力。针对企业用户，NextChat 提供专业版解决方案，具备品牌定制、细粒度权限控制、内部知识库整合及安全审计等功能，满足公司对数据隐私和个性化管理的高标准要求。",87618,"2026-04-05T07:20:52",[13,26],{"id":44,"name":45,"github_repo":46,"description_zh":47,"stars":48,"difficulty_score":23,"last_commit_at":49,"category_tags":50,"status":16},2268,"ML-For-Beginners","microsoft\u002FML-For-Beginners","ML-For-Beginners 是由微软推出的一套系统化机器学习入门课程，旨在帮助零基础用户轻松掌握经典机器学习知识。这套课程将学习路径规划为 12 周，包含 26 节精炼课程和 52 道配套测验，内容涵盖从基础概念到实际应用的完整流程，有效解决了初学者面对庞大知识体系时无从下手、缺乏结构化指导的痛点。\n\n无论是希望转型的开发者、需要补充算法背景的研究人员，还是对人工智能充满好奇的普通爱好者，都能从中受益。课程不仅提供了清晰的理论讲解，还强调动手实践，让用户在循序渐进中建立扎实的技能基础。其独特的亮点在于强大的多语言支持，通过自动化机制提供了包括简体中文在内的 50 多种语言版本，极大地降低了全球不同背景用户的学习门槛。此外，项目采用开源协作模式，社区活跃且内容持续更新，确保学习者能获取前沿且准确的技术资讯。如果你正寻找一条清晰、友好且专业的机器学习入门之路，ML-For-Beginners 将是理想的起点。",84991,"2026-04-05T10:45:23",[14,51,52,53,15,54,26,13,55],"数据工具","视频","插件","其他","音频",{"id":57,"name":58,"github_repo":59,"description_zh":60,"stars":61,"difficulty_score":10,"last_commit_at":62,"category_tags":63,"status":16},3128,"ragflow","infiniflow\u002Fragflow","RAGFlow 是一款领先的开源检索增强生成（RAG）引擎，旨在为大语言模型构建更精准、可靠的上下文层。它巧妙地将前沿的 RAG 技术与智能体（Agent）能力相结合，不仅支持从各类文档中高效提取知识，还能让模型基于这些知识进行逻辑推理和任务执行。\n\n在大模型应用中，幻觉问题和知识滞后是常见痛点。RAGFlow 通过深度解析复杂文档结构（如表格、图表及混合排版），显著提升了信息检索的准确度，从而有效减少模型“胡编乱造”的现象，确保回答既有据可依又具备时效性。其内置的智能体机制更进一步，使系统不仅能回答问题，还能自主规划步骤解决复杂问题。\n\n这款工具特别适合开发者、企业技术团队以及 AI 研究人员使用。无论是希望快速搭建私有知识库问答系统，还是致力于探索大模型在垂直领域落地的创新者，都能从中受益。RAGFlow 提供了可视化的工作流编排界面和灵活的 API 接口，既降低了非算法背景用户的上手门槛，也满足了专业开发者对系统深度定制的需求。作为基于 Apache 2.0 协议开源的项目，它正成为连接通用大模型与行业专有知识之间的重要桥梁。",77062,"2026-04-04T04:44:48",[15,14,13,26,54],{"id":65,"github_repo":66,"name":67,"description_en":68,"description_zh":69,"ai_summary_zh":69,"readme_en":70,"readme_zh":71,"quickstart_zh":72,"use_case_zh":73,"hero_image_url":74,"owner_login":75,"owner_name":75,"owner_avatar_url":76,"owner_bio":77,"owner_company":78,"owner_location":78,"owner_email":78,"owner_twitter":78,"owner_website":78,"owner_url":79,"languages":80,"stars":97,"forks":98,"last_commit_at":99,"license":100,"difficulty_score":23,"env_os":101,"env_gpu":102,"env_ram":103,"env_deps":104,"category_tags":112,"github_topics":113,"view_count":23,"oss_zip_url":78,"oss_zip_packed_at":78,"status":16,"created_at":121,"updated_at":122,"faqs":123,"releases":153},3544,"LAMDA-Tabular\u002FTALENT","TALENT","A comprehensive toolkit and benchmark for tabular data learning, featuring 35+ deep methods, more than 10 classical methods, and 300 diverse tabular datasets.","TALENT 是一个专为表格数据学习打造的综合工具箱与基准测试平台，旨在解决表格数据处理中模型选择困难、评估标准不一以及复现成本高等痛点。它集成了超过 35 种前沿深度学习方法和 10 余种经典算法，并提供了涵盖多领域、不同规模分布的 300 个多样化数据集，让用户能在统一框架下高效对比和验证模型性能。\n\n无论是刚入门的数据科学新手，还是深耕算法的研究人员或开发者，都能从 TALENT 中获益。它不仅内置了强大的数据预处理、归一化及编码功能，还支持灵活的超参数调优，极大地降低了实验门槛。其独特的技术亮点在于极高的可扩展性，用户可以轻松添加自定义数据集或新算法，同时支持多种评估指标，满足不同场景需求。此外，项目持续更新，不断纳入如 RFM、Real-TabPFN 等最新科研成果，确保用户始终能接触到领域内的最先进技术。通过提供标准化的实验环境与丰富的资源，TALENT 致力于推动表格数据深度学习领域的规范化发展与技术创新。","\u003Cp align=\"center\">\n\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FLAMDA-Tabular_TALENT_readme_3ae20edd7249.png\"  width=\"1000px\">\n\u003C\u002Fp>\n\u003Cp align=\"center\">\n    \u003Ca href='https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.04057'>\u003Cimg src='https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FArxiv-2407.04057-b31b1b.svg?logo=arXiv'>\u003C\u002Fa>\n    \u003Ca href='https:\u002F\u002Fzhuanlan.zhihu.com\u002Fp\u002F708721145'>\u003Cimg src='https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F中文解读-b0.svg?logo=zhihu'>\u003C\u002Fa>\n    \u003Ca href='https:\u002F\u002Fhuggingface.co\u002Fdatasets\u002FLAMDA-Tabular\u002FTALENT'>\u003Cimg src='https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F%F0%9F%A4%97-TALENT-green'>\u003C\u002Fa>\n  \u003Ca href=\"\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fqile2000\u002FLAMDA-TALENT?color=4fb5ee\">\u003C\u002Fa>\n  \u003Ca href=\"\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Flast-commit\u002Fqile2000\u002FLAMDA-TALENT?color=blue\">\u003C\u002Fa>\n   \u003Cbr>\n    \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPYTORCH-2.0.1-red?style=for-the-badge&logo=pytorch\" alt=\"PyTorch - Version\" height=\"21\">\n    \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPYTHON-3.10-red?style=for-the-badge&logo=python&logoColor=white\" alt=\"Python - Version\" height=\"21\">\n    \u003Ca href=\"\">\n    \u003Ca href='https:\u002F\u002Flamda-talent.readthedocs.io\u002Fen\u002Flatest\u002F?badge=latest'>\n    \u003Cimg src='https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FLAMDA-Tabular_TALENT_readme_13d664e1afd7.png' alt='Documentation Status' \u002F>\n\u003C\u002Fa>\u003Cimg src=\"https:\u002F\u002Fblack.readthedocs.io\u002Fen\u002Fstable\u002F_static\u002Flicense.svg\">\u003C\u002Fa>\n\u003C\u002Fp>\n\u003Cdiv align=\"center\">\n    \u003Cp>\n        TALENT: A Tabular Analytics and Learning Toolbox\n    \u003Cp>\n    \u003Cp>\n        \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.04057\">[Paper]\u003C\u002Fa> \u003Ca href=\"https:\u002F\u002Fzhuanlan.zhihu.com\u002Fp\u002F708721145\">[中文解读]\u003C\u002Fa> \u003Ca href=\"https:\u002F\u002Flamda-talent.readthedocs.io\u002Fen\u002Flatest\">[Docs]\u003C\u002Fa> \n    \u003Cp>\n\u003C\u002Fdiv>\n\n\n\n\n\n---\n\n## 🎉 Introduction\n\nWelcome to **TALENT**, a benchmark with a comprehensive machine learning toolbox designed to enhance model performance on tabular data. TALENT integrates advanced deep learning models, classical algorithms, and efficient hyperparameter tuning, offering robust preprocessing capabilities to optimize learning from tabular datasets. The toolbox is user-friendly and adaptable, catering to both novice and expert data scientists.\n\n**TALENT** offers the following advantages:\n\n- **Diverse Methods**: Includes various classical methods, tree-based methods, and the latest popular deep learning methods.\n- **Extensive Dataset Collection**: Equipped with 300 datasets, covering a wide range of task types, size distributions, and dataset domains.\n- **Customizability**: Easily allows the addition of datasets and methods.\n- **Versatile Support**: Supports diverse normalization, encoding, and metrics.\n\n## 📚Citing TALENT\n\n**If you use any content of this repo for your work, please cite the following bib entries:**\n\n```bibtex\n@article{ye2024closerlookdeeplearning,\n         title={A Closer Look at Deep Learning on Tabular Data}, \n         author={Han-Jia Ye and \n         \t\t Si-Yang Liu and \n         \t\t Hao-Run Cai and \n         \t\t Qi-Le Zhou and \n         \t\t De-Chuan Zhan},\n         journal={arXiv preprint arXiv:2407.00956},\n         year={2024}\n}\n\n@article{JMLR:v26:25-0512,\n  author  = {Si-Yang Liu and\n\t\t\t Hao-Run Cai and\n \t\t\t Qi-Le Zhou and\n\t\t\t Huai-Hong Yin and\n\t\t\t Tao Zhou and\n\t\t\t Jun-Peng Jiang and\n\t\t\t Han-Jia Ye},\n  title   = {Talent: A Tabular Analytics and Learning Toolbox},\n  journal = {Journal of Machine Learning Research},\n  year    = {2025},\n  volume  = {26},\n  number  = {226},\n  pages   = {1--16},\n  url     = {http:\u002F\u002Fjmlr.org\u002Fpapers\u002Fv26\u002F25-0512.html}\n}\n```\n\n\n\n\n## 📰 What's New\n\n- [2026-03]🌟 We have updated the TALENT-extension datasets and results. [Link](https:\u002F\u002Fbox.nju.edu.cn\u002Fd\u002Fb7b23a19ee054aaba7b6\u002F?p=%2F&mode=list)\n- [2025-11]🌟 Add [RFM](https:\u002F\u002Fwww.science.org\u002Fdoi\u002F10.1126\u002Fscience.adi5639) (Science).\n- [2025-11]🌟 Add [Real-TabPFN](https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.03971).\n- [2025-11]🌟 Add [LimiX](https:\u002F\u002Farxiv.org\u002Fabs\u002F2509.03505).\n- [2025-09]🌟 Add [xRFM](https:\u002F\u002Farxiv.org\u002Fabs\u002F2508.10053).\n- [2025-08]🌟 Add [Mitra](https:\u002F\u002Fwww.amazon.science\u002Fblog\u002Fmitra-mixed-synthetic-priors-for-enhancing-tabular-foundation-models).\n- [2025-06]🌟 Add [TabAutoPNPNet](https:\u002F\u002Fwww.mdpi.com\u002F2079-9292\u002F14\u002F6\u002F1165) (Electronics 2025).\n- [2025-06]🌟 Add [TabICL](https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.05564) (ICML 2025). The current code is based on TabICL v0.1.2.\n- [2025-05]🌟 Check out our three papers **[MMTU](https:\u002F\u002Fgithub.com\u002FLAMDA-Tabular\u002FMMTU)**, **[Tabular-Temporal-Shift](https:\u002F\u002Fgithub.com\u002FLAMDA-Tabular\u002FTabular-Temporal-Shift)**, and [**BETA**](https:\u002F\u002Fgithub.com\u002FLAMDA-Tabular\u002FBETA) accepted at ICML 2025!\n- [2025-04]🌟 Check out our new survey [Representation Learning for Tabular Data: A Comprehensive Survey](https:\u002F\u002Farxiv.org\u002Fabs\u002F2504.16109) ([Repo](https:\u002F\u002Fgithub.com\u002FLAMDA-Tabular\u002FTabular-Survey)). We organize existing methods into three main categories according to their generalization capabilities: specialized, transferable, and general models, which provides a comprehensive taxonomy for deep tabular representation methods.🚀🚀🚀\n- [2025-02]🌟 Add [T2Gformer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.16887) (AAAI 2023).\n- [2025-02]🌟 Add [TabPFN v2](https:\u002F\u002Fdoi.org\u002F10.1038\u002Fs41586-024-08328-6) (Nature).\n- [2025-02]🌟 Thanks to [Hengzhe Zhang](https:\u002F\u002Fhengzhe-zhang.github.io\u002F) for providing a [Scikit-Learn compatible wrapper](https:\u002F\u002Fgithub.com\u002Fhengzhe-zhang\u002Fscikit-TALENT) for TALENT!\n- [2025-01]🌟 Check out our new baseline [ModernNCA](https:\u002F\u002Fopenreview.net\u002Fpdf?id=JytL2MrlLT) (**ICLR 2025**), inspired by traditional **Neighbor Component Analysis**, which outperforms both tree-based and other deep tabular models, while also reducing training time and model size!🚀🚀🚀\n- [2025-01]🌟 Check out our [latest version of the benchmark paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.00956) for updated and expanded results and analysis!\n- [2025-01]🌟We have curated and released [new benchmark datasets](https:\u002F\u002Fdrive.google.com\u002Fdrive\u002Ffolders\u002F1j1zt3zQIo8dO6vkO-K-WE6pSrl71bf0z?usp=drive_link), along with updated [results](https:\u002F\u002F6sy666.github.io\u002FTALENT-Results\u002F) of the dataset across a broader range of methods. This update focuses on enhancing dataset quality, including removing duplicates, and correcting tasks where bin-class was mistakenly treated as regression. We have also separated the larger datasets and formed the basic benchmark (300 datasets, including 120 bin-class, 80 multi-class, and 100 regression), and the large benchmark (22 datasets).\n- [2024-12]🌟 Add [TabM](https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.24210) (ICLR 2025).\n- [2024-09]🌟 Add [Trompt](https:\u002F\u002Farxiv.org\u002Fabs\u002F2305.18446) (ICML 2023).\n- [2024-09]🌟 Add [AMFormer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.02334) (AAAI 2024).\n- [2024-08]🌟 Add [GRANDE](https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.17130) (ICLR 2024).\n- [2024-08]🌟 Add [Excelformer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2301.02819) (KDD 2024).\n- [2024-08]🌟 Add [MLP_PLR](https:\u002F\u002Farxiv.org\u002Fabs\u002F2203.05556) (NeurIPS 2022).\n- [2024-07]🌟 Add [RealMLP](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.04491)(NeurIPS 2024).\n- [2024-07]🌟 Add [ProtoGate](https:\u002F\u002Farxiv.org\u002Fabs\u002F2306.12330) (ICML 2024).\n- [2024-07]🌟 Add [BiSHop](https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.03830) (ICML 2024).\n- [2024-06]🌟 Check out our new baseline [ModernNCA](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.03257), inspired by traditional **Neighbor Component Analysis**, which outperforms both tree-based and other deep tabular models, while also reducing training time and model size!\n- [2024-06]🌟 Check out our [benchmark paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.00956) about tabular data, which provides comprehensive evaluations of classical and deep tabular methods based on our toolbox in a fair manner!\n\n## 🌟 Methods\n\nTALENT integrates an extensive array of 30+ deep learning architectures for tabular data, including but not limited to:\n\n1. **MLP**: A multi-layer neural network, which is implemented according to [RTDL](https:\u002F\u002Farxiv.org\u002Fabs\u002F2106.11959).\n2. **ResNet**: A DNN that uses skip connections across many layers, which is implemented according to [RTDL](https:\u002F\u002Farxiv.org\u002Fabs\u002F2106.11959).\n3. **[SNN](https:\u002F\u002Farxiv.org\u002Fabs\u002F1706.02515)**: An MLP-like architecture utilizing the SELU activation, which facilitates the training of deeper neural networks.\n4. **[DANets](https:\u002F\u002Farxiv.org\u002Fabs\u002F2112.02962)**: A neural network designed to enhance tabular data processing by grouping correlated features and reducing computational complexity.\n5. **[TabCaps](https:\u002F\u002Fopenreview.net\u002Fpdf?id=OgbtSLESnI)**: A capsule network that encapsulates all feature values of a record into vectorial features.\n6. **[DCNv2](https:\u002F\u002Farxiv.org\u002Fabs\u002F2008.13535)**: Consists of an MLP-like module combined with a feature crossing module, which includes both linear layers and multiplications.\n7. **[NODE](https:\u002F\u002Farxiv.org\u002Fabs\u002F1909.06312)**: A tree-mimic method that generalizes oblivious decision trees, combining gradient-based optimization with hierarchical representation learning.\n8. **[GrowNet](https:\u002F\u002Farxiv.org\u002Fabs\u002F2002.07971)**: A gradient boosting framework that uses shallow neural networks as weak learners.\n9. **[TabNet](https:\u002F\u002Farxiv.org\u002Fabs\u002F1908.07442)**: A tree-mimic method using sequential attention for feature selection, offering interpretability and self-supervised learning capabilities.\n10. **[TabR](https:\u002F\u002Farxiv.org\u002Fabs\u002F2307.14338)**: A deep learning model that integrates a KNN component to enhance tabular data predictions through an efficient attention-like mechanism.\n11. **[ModernNCA](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.03257)**: A deep tabular model inspired by traditional Neighbor Component Analysis, which makes predictions based on the relationships with neighbors in a learned embedding space.\n12. **[DNNR](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.08434)**: Enhances KNN by using local gradients and Taylor approximations for more accurate and interpretable predictions.\n13. **[AutoInt](https:\u002F\u002Farxiv.org\u002Fabs\u002F1810.11921)**: A token-based method that uses a multi-head self-attentive neural network to automatically learn high-order feature interactions.\n14. **[Saint](https:\u002F\u002Farxiv.org\u002Fabs\u002F2106.01342)**: A token-based method that leverages row and column attention mechanisms for tabular data.\n15. **[TabTransformer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2012.06678)**: A token-based method that enhances tabular data modeling by transforming categorical features into contextual embeddings.\n16. **[FT-Transformer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2106.11959)**: A token-based method which transforms features to embeddings and applies a series of attention-based transformations to the embeddings.\n17. **[TANGOS](https:\u002F\u002Fopenreview.net\u002Fpdf?id=n6H86gW8u0d)**: A regularization-based method for tabular data that uses gradient attributions to encourage neuron specialization and orthogonalization.\n18. **[SwitchTab](https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.02013)**: A self-supervised method tailored for tabular data that improves representation learning through an asymmetric encoder-decoder framework. Following the original paper, our toolkit uses a supervised learning form, optimizing both reconstruction and supervised loss in each epoch.\n19. **[PTaRL](https:\u002F\u002Fopenreview.net\u002Fpdf?id=G32oY4Vnm8)**: A regularization-based framework that enhances prediction by constructing and projecting into a prototype-based space.\n20. **[TabPFN](https:\u002F\u002Farxiv.org\u002Fabs\u002F2207.01848)**: A general model which involves the use of pre-trained deep neural networks that can be directly applied to any tabular task.\n21. **[HyperFast](https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.14335)**: A meta-trained hypernetwork that generates task-specific neural networks for instant classification of tabular data.\n22. **[TabPTM](https:\u002F\u002Farxiv.org\u002Fabs\u002F2311.00055)**: A general method for tabular data that standardizes heterogeneous datasets using meta-representations, allowing a pre-trained model to generalize to unseen datasets without additional training.\n23. **[BiSHop](https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.03830)**: An end-to-end framework for deep tabular learning which leverages a sparse Hopfield model with adaptable sparsity, enhanced by column-wise and row-wise modules.\n24. **[ProtoGate](https:\u002F\u002Farxiv.org\u002Fabs\u002F2306.12330)**: A prototype-based model for feature selection in HDLSS biomedical data that adapts global and local feature selection to enhance prediction accuracy and interpretability, addressing co-adaptation issues through a non-parametric prototype-based mechanism.\n25. **[RealMLP](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.04491)**: An improved multilayer perceptron (MLP).\n26. **[MLP_PLR](https:\u002F\u002Farxiv.org\u002Fabs\u002F2203.05556)**: An improved multilayer perceptron (MLP), which utilizes periodic activations.\n27. **[Excelformer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2301.02819)**: A deep learning model for tabular data prediction, featuring a semi-permeable attention module to address rotational invariance, tailored data augmentation, and an attentive feedforward network, making it a reliable solution across diverse datasets.\n28. **[GRANDE](https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.17130)**: A tree-mimic method for learning hard, axis-aligned decision tree ensembles using end-to-end gradient descent.\n29. **[AMFormer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.02334)**: A token-based method which improves the transformer architecture for tabular data by incorporating parallel addition and multiplication attention mechanisms, utilizing prompt tokens to constrain feature interactions.\n30. **[Trompt](https:\u002F\u002Farxiv.org\u002Fabs\u002F2305.18446)**: A prompt-based deep neural network for tabular data that separates learning into intrinsic column features and sample-specific feature importance.\n31. **[TabM](https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.24210)** :  A model based on MLP and variations of BatchEnsemble.\n32. **[TabPFN v2](https:\u002F\u002Fdoi.org\u002F10.1038\u002Fs41586-024-08328-6)**: A general model which involves the use of pre-trained deep neural networks that can be directly applied to any tabular task.\n33. **[T2Gformer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.16887)**: A Transformer network for tabular learning that processes data guided by relation graphs and uses a Cross-level Readout for global semantics in prediction.\n34. **[TabICL](https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.05564)**: A comparable tabular foundation model with performance on par with TabPFN v2.\n35. **[TabAutoPNPNet](https:\u002F\u002Fwww.mdpi.com\u002F2079-9292\u002F14\u002F6\u002F1165)**: A tabular model based on periodicity, particularly the Fourier transform and Chebyshev polynomials, with performance on par with or superior to FT-Transformer. \n36. **[Mitra](https:\u002F\u002Fwww.amazon.science\u002Fblog\u002Fmitra-mixed-synthetic-priors-for-enhancing-tabular-foundation-models)**: A tabular foundation model learned with mixed synthetic priors.\n37. **[LimiX](https:\u002F\u002Farxiv.org\u002Fabs\u002F2509.03505)**: A tabular foundation model that leverages transformers to support a wide range of tasks, from prediction to imputation and causal inference, within a unified architecture.\n38. **[Real-TabPFN](https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.03971)**: An enhanced tabular foundation model that extends TabPFNv2 through continued pre-training on real-world datasets for classification tasks. \n39. **[RFM](https:\u002F\u002Fwww.science.org\u002Fdoi\u002F10.1126\u002Fscience.adi5639)**: A non-deep, backpropagation-free feature learning algorithm, iteratively applies AGOP to a kernel machine to adaptively learn task-specific features.\n40. **[xRFM](https:\u002F\u002Farxiv.org\u002Fabs\u002F2508.10053)**: A tabular model that combines RFMs with an adaptive tree structure, enabling it to learn features local to data subsets and scale log-linearly with the number of samples.\n\n\n🔧 If you want to check the **default hyperparameters and hyperparameter search spaces** of all methods, please visit:  \n👉 [https:\u002F\u002F6sy666.github.io\u002FTALENT-Configs\u002F](https:\u002F\u002F6sy666.github.io\u002FTALENT-Configs\u002F)\n\n## ☄️ How to Use TALENT\n\n### 🕹️ Quick Start\n\nInstall with the newest version through GitHub:\n\n```bash\n$ pip install git+https:\u002F\u002Fgithub.com\u002FLAMDA-Tabular\u002FTALENT.git@main --upgrade\n```\n\nTry a demo `train_model_deep.py` :\n\n```python\n\nfrom tqdm import tqdm\nfrom TALENT.model.utils import get_deep_args,show_results,tune_hyper_parameters,get_method,set_seeds\nfrom TALENT.model.lib.data import get_dataset\n\nif __name__ == '__main__':\n    loss_list, results_list, time_list = [], [], []\n    args,default_para,opt_space = get_deep_args()\n    train_val_data,test_data,info = get_dataset(args.dataset,args.dataset_path)\n    if args.tune:\n        args = tune_hyper_parameters(args,opt_space,train_val_data,info)\n    for seed in tqdm(range(args.seed_num)):\n        args.seed = seed    # update seed  \n        set_seeds(args.seed)\n        method = get_method(args.model_type)(args, info['task_type'] == 'regression')\n        time_cost = method.fit(train_val_data, info)    \n        vl, vres, metric_name, predict_logits = method.predict(test_data, info, model_name=args.evaluate_option)\n\t    loss_list.append(vl)\n        results_list.append(vres)\n        time_list.append(time_cost)\n\n    show_results(args,info, metric_name,loss_list,results_list,time_list)\n\n```\n\n\n\n```bash\npython train_model_deep.py --model_type MODEL_NAME\n```\n\n\n\n> For researchers:\n\n### 🕹️ Clone\n\nClone this GitHub repository:\n\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002FLAMDA-Tabular\u002FTALENT\ncd TALENT\u002Ftest\n```\n\n### 🔑 Run experiment\n\n1. Edit the `configs\u002Fdefault\u002F[MODEL_NAME].json`  and `config\u002Fopt_space\u002F[MODEL_NAME].json` for global settings and hyperparameters.\n\n2. Run:\n\n    ```bash\n    python train_model_deep.py --model_type MODEL_NAME\n    ```\n    for deep methods, or:\n    ```bash\n    python train_model_classical.py --model_type MODEL_NAME\n    ```\n    for classical methods.\t\n\n### 🛠️ How to Add New Methods\n\nFor methods like the MLP class that only need to design the model, you only need to:\n\n- Add the model class in `model\u002Fmodels`.\n- Inherit from `model\u002Fmethods\u002Fbase.py` and override the `construct_model()` method in the new class.\n- Add the method name in the `get_method` function in `model\u002Futils.py`.\n- Add the parameter settings for the new method in `configs\u002Fdefault\u002F[MODEL_NAME].json` and `configs\u002Fopt_space\u002F[MODEL_NAME].json`.\n\nFor other methods that require changing the training process, partially override functions based on `model\u002Fmethods\u002Fbase.py`. For details, refer to the implementation of other methods in `model\u002Fmethods\u002F`.\n\nSee our [Contribution Guide](CONTRIBUTING.md) for more details.\n\n### 📦 Dependencies\n\n```bash\n   pip install -r requirements.txt\n```\n\n\nIf you want to use **TabR**, you have to manually install faiss, which is only available on **conda**:\n\n```bash\nconda install faiss-gpu -c pytorch\n```\n\n## 🗂️ Benchmark Datasets\n\nDatasets are available at [Google Drive](https:\u002F\u002Fdrive.google.com\u002Fdrive\u002Ffolders\u002F1j1zt3zQIo8dO6vkO-K-WE6pSrl71bf0z?usp=drive_link).\n\n### 📂 How to Place Datasets\n\nDatasets are placed in the project's current directory, corresponding to the file name specified by `args.dataset_path`. For instance, if the project is `LAMDA-TALENT`, the data should be placed in `LAMDA-TALENT\u002Fargs.dataset_path\u002Fargs.dataset`.\n\nEach dataset folder `args.dataset` consists of:\n\n- Numeric features: `N_train\u002Fval\u002Ftest.npy` (can be omitted if there are no numeric features)\n\n- Categorical features: `C_train\u002Fval\u002Ftest.npy` (can be omitted if there are no categorical features)\n\n- Labels: `y_train\u002Fval\u002Ftest.npy`\n\n- `info.json`, which must include the following three contents (task_type can be \"regression\", \"multiclass\" or \"binclass\"):\n  \n  ```json\n  {\n    \"task_type\": \"regression\", \n    \"n_num_features\": 10,\n    \"n_cat_features\": 10\n  }\n  ```\n\n## 📝 Experimental Results\n\nWe provide comprehensive evaluations of classical and deep tabular methods based on our toolbox in a fair manner in the Figure. Three tabular prediction tasks, namely, binary classification, multi-class classification, and regression, are considered, and each subfigure represents a different task type.\n\nWe use `Accuracy` and `RMSE` as the metrics for classification tasks and regression tasks, respectively. To calibrate the metrics, we choose the average performance rank to compare all methods, where a lower rank indicates better performance, following  [Sheskin (2003)](https:\u002F\u002Fwww.taylorfrancis.com\u002Fbooks\u002Fmono\u002F10.1201\u002F9781420036268\u002Fhandbook-parametric-nonparametric-statistical-procedures-david-sheskin). Efficiency is calculated by the average training time in seconds, with lower values denoting better time efficiency. The model size is visually indicated by the radius of the circles, offering a quick glance at the trade-off between model complexity and performance.\n\n\u003C!-- The classical method `SVM` provided in TALENT is a `LinearSVM` to ensure faster training.  We also consider the `Dummy` baseline, which outputs the label of the major class and the average labels for classification and regression tasks, respectively. -->\n\n\u003Cdiv align=\"center\" style=\"text-align:center;\">\n\n  \u003Cdiv style=\"display:inline-block; width:45%; vertical-align:top; margin-bottom:5px;\">\n    \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FLAMDA-Tabular_TALENT_readme_2ebfaaa192f5.png\" alt=\"Binary classification\" style=\"width:45%; border-radius:8px;\" \u002F>\n    \u003Cdiv>(a) Binary classification\u003C\u002Fdiv>\n  \u003C\u002Fdiv>\u003Cdiv style=\"display:inline-block; width:45%; vertical-align:top; margin-bottom:5px;\">\n    \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FLAMDA-Tabular_TALENT_readme_da07af822efc.png\" alt=\"Multiclass Classification\" style=\"width:45%; border-radius:8px;\" \u002F>\n    \u003Cdiv>(b) Multiclass classification\u003C\u002Fdiv>\n  \u003C\u002Fdiv>\n\n  \u003Cdiv style=\"display:inline-block; width:45%; vertical-align:top; margin-bottom:5px;\">\n    \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FLAMDA-Tabular_TALENT_readme_97e6e7e0e661.png\" alt=\"Regression\" style=\"width:45%; border-radius:8px;\" \u002F>\n    \u003Cdiv>(c) Regression\u003C\u002Fdiv>\n  \u003C\u002Fdiv>\u003Cdiv style=\"display:inline-block; width:45%; vertical-align:top; margin-bottom:5px;\">\n    \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FLAMDA-Tabular_TALENT_readme_5ce9070ebd8c.png\" alt=\"All tasks\" style=\"width:45%; border-radius:8px;\" \u002F>\n    \u003Cdiv>(d) All tasks\u003C\u002Fdiv>\n  \u003C\u002Fdiv>\n\n\u003C\u002Fdiv>\n\n\u003C!-- From the comparison, we observe that **CatBoost** achieves the best average rank in most classification and regression tasks. Among all deep tabular methods, **ModernNCA** performs the best in most cases while maintaining an acceptable training cost. These results highlight the effectiveness of CatBoost and ModernNCA in handling various tabular prediction tasks, making them suitable choices for practitioners seeking high performance and efficiency.\n\nThese visualizations serve as an effective tool for quickly and fairly assessing the strengths and weaknesses of various tabular methods across different task types, enabling researchers and practitioners to make informed decisions when selecting suitable modeling techniques for their specific needs. -->\n\n## 👨🏫 Acknowledgments\n\nWe thank the following repos for providing helpful components\u002Ffunctions in our work:\n\n- [Rtdl-revisiting-models](https:\u002F\u002Fgithub.com\u002Fyandex-research\u002Frtdl-revisiting-models)\n- [Rtdl-num-embeddings](https:\u002F\u002Fgithub.com\u002Fyandex-research\u002Frtdl-num-embeddings)\n- [Tabular-dl-tabr](https:\u002F\u002Fgithub.com\u002Fyandex-research\u002Ftabular-dl-tabr)\n- [DANet](https:\u002F\u002Fgithub.com\u002FWhatAShot\u002FDANet)\n- [TabCaps](https:\u002F\u002Fgithub.com\u002FWhatAShot\u002FTabCaps)\n- [DNNR](https:\u002F\u002Fgithub.com\u002Fyounader\u002Fdnnr)\n- [PTaRL](https:\u002F\u002Fgithub.com\u002FHangtingYe\u002FPTaRL)\n- [Saint](https:\u002F\u002Fgithub.com\u002Fsomepago\u002Fsaint)\n- [SwitchTab](https:\u002F\u002Fgithub.com\u002Favivnur\u002FSwitchTab)\n- [TabNet](https:\u002F\u002Fgithub.com\u002Fdreamquark-ai\u002Ftabnet)\n- [TabPFN](https:\u002F\u002Fgithub.com\u002Fautoml\u002FTabPFN)\n- [Tabtransformer-pytorch](https:\u002F\u002Fgithub.com\u002Flucidrains\u002Ftab-transformer-pytorch)\n- [TANGOS](https:\u002F\u002Fgithub.com\u002Falanjeffares\u002FTANGOS)\n- [GrowNet](https:\u002F\u002Fgithub.com\u002Fsbadirli\u002FGrowNet)\n- [HyperFast](https:\u002F\u002Fgithub.com\u002FAI-sandbox\u002FHyperFast)\n- [BiSHop](https:\u002F\u002Fgithub.com\u002FMAGICS-LAB\u002FBiSHop)\n- [ProtoGate](https:\u002F\u002Fgithub.com\u002FSilenceX12138\u002FProtoGate)\n- [Pytabkit](https:\u002F\u002Fgithub.com\u002Fdholzmueller\u002Fpytabkit)\n- [Excelformer](https:\u002F\u002Fgithub.com\u002FWhatAShot\u002FExcelFormer)\n- [GRANDE](https:\u002F\u002Fgithub.com\u002Fs-marton\u002FGRANDE)\n- [AMFormer](https:\u002F\u002Fgithub.com\u002Faigc-apps\u002FAMFormer)\n- [TabM](https:\u002F\u002Fgithub.com\u002Fyandex-research\u002Ftabm)\n- [TabICL](https:\u002F\u002Fgithub.com\u002Fsoda-inria\u002Ftabicl)\n- [TabAutoPNPNet](https:\u002F\u002Fgithub.com\u002Fmatteo-rizzo\u002Fperiodic-tabular-dl)\n- [LimiX](https:\u002F\u002Fgithub.com\u002Flimix-ldm\u002FLimiX)\n- [xRFM](https:\u002F\u002Fgithub.com\u002Fdmbeaglehole\u002FxRFM)\n\n## 🤗 Contact\n\nIf there are any questions, please feel free to propose new features by opening an issue or contact the author: **Si-Yang Liu** ([liusy@lamda.nju.edu.cn](mailto:liusy@lamda.nju.edu.cn)) and **Hao-Run Cai** ([caihr@lamda.nju.edu.cn](mailto:caihr@lamda.nju.edu.cn)) and **Qile Zhou** ([zhouql@lamda.nju.edu.cn](mailto:zhouql@lamda.nju.edu.cn)) and **Jun-Peng Jiang** ([jiangjp@lamda.nju.edu.cn](mailto:jiangjp@lamda.nju.edu.cn)) and **Huai-Hong Yin** ([yinhh@lamda.nju.edu.cn](mailto:yinhh@lamda.nju.edu.cn)) and **Tao Zhou** ([zhout@lamda.nju.edu.cn]) and **Han-Jia Ye** ([yehj@lamda.nju.edu.cn](mailto:yehj@lamda.nju.edu.cn)). Enjoy the code.\n\n## 🚀 Star History\n\n[![Star History Chart](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FLAMDA-Tabular_TALENT_readme_5ac421800874.png)](https:\u002F\u002Fstar-history.com\u002F#LAMDA-Tabular\u002FTALENT&Date)\n\n> Thanks [LAMDA-PILOT](https:\u002F\u002Fgithub.com\u002FLAMDA-CL\u002FLAMDA-PILOT) and [LAMDA-ZhiJian](https:\u002F\u002Fgithub.com\u002Fzhangyikaii\u002FLAMDA-ZhiJian) for the template.\n","\u003Cp align=\"center\">\n\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FLAMDA-Tabular_TALENT_readme_3ae20edd7249.png\"  width=\"1000px\">\n\u003C\u002Fp>\n\u003Cp align=\"center\">\n    \u003Ca href='https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.04057'>\u003Cimg src='https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FArxiv-2407.04057-b31b1b.svg?logo=arXiv'>\u003C\u002Fa>\n    \u003Ca href='https:\u002F\u002Fzhuanlan.zhihu.com\u002Fp\u002F708721145'>\u003Cimg src='https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F中文解读-b0.svg?logo=zhihu'>\u003C\u002Fa>\n    \u003Ca href='https:\u002F\u002Fhuggingface.co\u002Fdatasets\u002FLAMDA-Tabular\u002FTALENT'>\u003Cimg src='https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F%F0%9F%A4%97-TALENT-green'>\u003C\u002Fa>\n  \u003Ca href=\"\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fqile2000\u002FLAMDA-TALENT?color=4fb5ee\">\u003C\u002Fa>\n  \u003Ca href=\"\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Flast-commit\u002Fqile2000\u002FLAMDA-TALENT?color=blue\">\u003C\u002Fa>\n   \u003Cbr>\n    \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPYTORCH-2.0.1-red?style=for-the-badge&logo=pytorch\" alt=\"PyTorch - Version\" height=\"21\">\n    \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPYTHON-3.10-red?style=for-the-badge&logo=python&logoColor=white\" alt=\"Python - Version\" height=\"21\">\n    \u003Ca href=\"\">\n    \u003Ca href='https:\u002F\u002Flamda-talent.readthedocs.io\u002Fen\u002Flatest\u002F?badge=latest'>\n    \u003Cimg src='https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FLAMDA-Tabular_TALENT_readme_13d664e1afd7.png' alt='Documentation Status' \u002F>\n\u003C\u002Fa>\u003Cimg src=\"https:\u002F\u002Fblack.readthedocs.io\u002Fen\u002Fstable\u002F_static\u002Flicense.svg\">\u003C\u002Fa>\n\u003C\u002Fp>\n\u003Cdiv align=\"center\">\n    \u003Cp>\n        TALENT：一个表格数据分析与学习工具箱\n    \u003Cp>\n    \u003Cp>\n        \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.04057\">[论文]\u003C\u002Fa> \u003Ca href=\"https:\u002F\u002Fzhuanlan.zhihu.com\u002Fp\u002F708721145\">[中文解读]\u003C\u002Fa> \u003Ca href=\"https:\u002F\u002Flamda-talent.readthedocs.io\u002Fen\u002Flatest\">[文档]\u003C\u002Fa> \n    \u003Cp>\n\u003C\u002Fdiv>\n\n\n\n\n\n---\n\n## 🎉 简介\n\n欢迎使用 **TALENT**，这是一个专为提升表格数据上模型性能而设计的、包含全面机器学习工具集的基准测试平台。TALENT 集成了先进的深度学习模型、经典算法以及高效的超参数调优方法，并提供强大的预处理功能，以优化从表格数据集中进行学习的效果。该工具箱易于使用且高度可扩展，能够满足初学者和资深数据科学家的需求。\n\n**TALENT** 具有以下优势：\n\n- **方法多样**：涵盖多种经典方法、基于树的方法以及最新的热门深度学习方法。\n- **丰富的数据集集合**：配备 300 个数据集，覆盖广泛的任务类型、规模分布及数据领域。\n- **可定制性**：可轻松添加新的数据集和方法。\n- **多功能支持**：支持多种归一化、编码和评估指标。\n\n## 📚 引用 TALENT\n\n**如果您在工作中使用了本仓库中的任何内容，请引用以下 BibTeX 条目：**\n\n```bibtex\n@article{ye2024closerlookdeeplearning,\n         title={A Closer Look at Deep Learning on Tabular Data}, \n         author={Han-Jia Ye and \n         \t\t Si-Yang Liu and \n         \t\t Hao-Run Cai and \n         \t\t Qi-Le Zhou and \n         \t\t De-Chuan Zhan},\n         journal={arXiv preprint arXiv:2407.00956},\n         year={2024}\n}\n\n@article{JMLR:v26:25-0512,\n  author  = {Si-Yang Liu and\n\t\t\t Hao-Run Cai and\n \t\t\t Qi-Le Zhou and\n\t\t\t Huai-Hong Yin and\n\t\t\t Tao Zhou and\n\t\t\t Jun-Peng Jiang and\n\t\t\t Han-Jia Ye},\n  title   = {Talent: A Tabular Analytics and Learning Toolbox},\n  journal = {Journal of Machine Learning Research},\n  year    = {2025},\n  volume  = {26},\n  number  = {226},\n  pages   = {1--16},\n  url     = {http:\u002F\u002Fjmlr.org\u002Fpapers\u002Fv26\u002F25-0512.html}\n}\n```\n\n## 📰 最新动态\n\n- [2026-03]🌟 我们更新了TALENT扩展数据集及结果。[链接](https:\u002F\u002Fbox.nju.edu.cn\u002Fd\u002Fb7b23a19ee054aaba7b6\u002F?p=%2F&mode=list)\n- [2025-11]🌟 新增[RFM](https:\u002F\u002Fwww.science.org\u002Fdoi\u002F10.1126\u002Fscience.adi5639)（Science）。\n- [2025-11]🌟 新增[Real-TabPFN](https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.03971)。\n- [2025-11]🌟 新增[LimiX](https:\u002F\u002Farxiv.org\u002Fabs\u002F2509.03505)。\n- [2025-09]🌟 新增[xRFM](https:\u002F\u002Farxiv.org\u002Fabs\u002F2508.10053)。\n- [2025-08]🌟 新增[Mitra](https:\u002F\u002Fwww.amazon.science\u002Fblog\u002Fmitra-mixed-synthetic-priors-for-enhancing-tabular-foundation-models)。\n- [2025-06]🌟 新增[TabAutoPNPNet](https:\u002F\u002Fwww.mdpi.com\u002F2079-9292\u002F14\u002F6\u002F1165)（Electronics 2025）。\n- [2025-06]🌟 新增[TabICL](https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.05564)（ICML 2025）。当前代码基于TabICL v0.1.2。\n- [2025-05]🌟 欢迎查看我们三篇被ICML 2025接收的论文：**[MMTU](https:\u002F\u002Fgithub.com\u002FLAMDA-Tabular\u002FMMTU)**、**[Tabular-Temporal-Shift](https:\u002F\u002Fgithub.com\u002FLAMDA-Tabular\u002FTabular-Temporal-Shift)** 和 **[BETA](https:\u002F\u002Fgithub.com\u002FLAMDA-Tabular\u002FBETA)**！\n- [2025-04]🌟 欢迎查看我们的新综述《表格数据表示学习：全面综述》（https:\u002F\u002Farxiv.org\u002Fabs\u002F2504.16109）（[项目仓库](https:\u002F\u002Fgithub.com\u002FLAMDA-Tabular\u002FTabular-Survey)）。我们根据模型的泛化能力，将现有方法分为三大类：专用模型、可迁移模型和通用模型，为深度表格表示方法提供了一个全面的分类体系。🚀🚀🚀\n- [2025-02]🌟 新增[T2Gformer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.16887)（AAAI 2023）。\n- [2025-02]🌟 新增[TabPFN v2](https:\u002F\u002Fdoi.org\u002F10.1038\u002Fs41586-024-08328-6)（Nature）。\n- [2025-02]🌟 感谢[Hengzhe Zhang](https:\u002F\u002Fhengzhe-zhang.github.io\u002F)为TALENT提供了[与Scikit-Learn兼容的封装库](https:\u002F\u002Fgithub.com\u002Fhengzhe-zhang\u002Fscikit-TALENT)！\n- [2025-01]🌟 欢迎查看我们的新基线[ModernNCA](https:\u002F\u002Fopenreview.net\u002Fpdf?id=JytL2MrlLT)（**ICLR 2025**），其灵感源自传统的**邻域成分分析**，在性能上超越了基于树的模型及其他深度表格模型，同时还能减少训练时间和模型规模！🚀🚀🚀\n- [2025-01]🌟 欢迎查看我们的[基准论文最新版本](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.00956)，其中包含了更新和扩充的结果与分析！\n- [2025-01]🌟 我们整理并发布了[新的基准数据集](https:\u002F\u002Fdrive.google.com\u002Fdrive\u002Ffolders\u002F1j1zt3zQIo8dO6vkO-K-WE6pSrl71bf0z?usp=drive_link)，同时更新了该数据集在更广泛方法上的[结果](https:\u002F\u002F6sy666.github.io\u002FTALENT-Results\u002F)。此次更新重点提升了数据集质量，包括去除重复样本，并修正了将二分类任务误认为回归任务的情况。我们还对大型数据集进行了拆分，形成了基础基准（300个数据集，包括120个二分类、80个多分类和100个回归任务）以及大型基准（22个数据集）。\n- [2024-12]🌟 新增[TabM](https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.24210)（ICLR 2025）。\n- [2024-09]🌟 新增[Trompt](https:\u002F\u002Farxiv.org\u002Fabs\u002F2305.18446)（ICML 2023）。\n- [2024-09]🌟 新增[AMFormer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.02334)（AAAI 2024）。\n- [2024-08]🌟 新增[GRANDE](https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.17130)（ICLR 2024）。\n- [2024-08]🌟 新增[Excelformer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2301.02819)（KDD 2024）。\n- [2024-08]🌟 新增[MLP_PLR](https:\u002F\u002Farxiv.org\u002Fabs\u002F2203.05556)（NeurIPS 2022）。\n- [2024-07]🌟 新增[RealMLP](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.04491)（NeurIPS 2024）。\n- [2024-07]🌟 新增[ProtoGate](https:\u002F\u002Farxiv.org\u002Fabs\u002F2306.12330)（ICML 2024）。\n- [2024-07]🌟 新增[BiSHop](https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.03830)（ICML 2024）。\n- [2024-06]🌟 欢迎查看我们的新基线[ModernNCA](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.03257)，其灵感源自传统的**邻域成分分析**，在性能上超越了基于树的模型及其他深度表格模型，同时还能减少训练时间和模型规模！\n- [2024-06]🌟 欢迎查看我们的关于表格数据的[基准论文](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.00956)，该论文基于我们的工具箱，以公平的方式对经典及深度表格方法进行了全面评估！\n\n## 🌟 方法\n\nTALENT整合了30余种针对表格数据的深度学习架构，其中包括但不限于：\n\n1. **MLP**：多层神经网络，根据[RTDL](https:\u002F\u002Farxiv.org\u002Fabs\u002F2106.11959)实现。\n2. **ResNet**：一种在多层之间使用跳跃连接的深度神经网络，同样根据[RTDL](https:\u002F\u002Farxiv.org\u002Fabs\u002F2106.11959)实现。\n3. **[SNN](https:\u002F\u002Farxiv.org\u002Fabs\u002F1706.02515)**：一种类似MLP的架构，采用SELU激活函数，有助于训练更深的神经网络。\n4. **[DANets](https:\u002F\u002Farxiv.org\u002Fabs\u002F2112.02962)**：一种旨在通过将相关特征分组并降低计算复杂度来增强表格数据处理能力的神经网络。\n5. **[TabCaps](https:\u002F\u002Fopenreview.net\u002Fpdf?id=OgbtSLESnI)**：一种胶囊网络，将一条记录的所有特征值封装为向量特征。\n6. **[DCNv2](https:\u002F\u002Farxiv.org\u002Fabs\u002F2008.13535)**：由一个类似MLP的模块与特征交叉模块组成，该模块包含线性层和乘法操作。\n7. **[NODE](https:\u002F\u002Farxiv.org\u002Fabs\u002F1909.06312)**：一种模仿树结构的方法，推广了无记忆决策树，结合基于梯度的优化与层次化表示学习。\n8. **[GrowNet](https:\u002F\u002Farxiv.org\u002Fabs\u002F2002.07971)**：一个梯度提升框架，使用浅层神经网络作为弱学习器。\n9. **[TabNet](https:\u002F\u002Farxiv.org\u002Fabs\u002F1908.07442)**：一种利用序列注意力进行特征选择的树状模仿方法，具有可解释性和自监督学习能力。\n10. **[TabR](https:\u002F\u002Farxiv.org\u002Fabs\u002F2307.14338)**：一种深度学习模型，集成KNN组件，通过高效的类注意力机制提升表格数据预测性能。\n11. **[ModernNCA](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.03257)**：受传统邻域成分分析启发的深度表格模型，基于学习到的嵌入空间中与邻居的关系进行预测。\n12. **[DNNR](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.08434)**：通过使用局部梯度和泰勒近似，提升KNN的准确性和可解释性。\n13. **[AutoInt](https:\u002F\u002Farxiv.org\u002Fabs\u002F1810.11921)**：一种基于标记的方法，利用多头自注意力神经网络自动学习高阶特征交互。\n14. **[Saint](https:\u002F\u002Farxiv.org\u002Fabs\u002F2106.01342)**：一种基于标记的方法，利用行和列注意力机制处理表格数据。\n15. **[TabTransformer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2012.06678)**：一种基于标记的方法，通过将类别特征转换为上下文嵌入来增强表格数据建模能力。\n16. **[FT-Transformer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2106.11959)**：一种基于标记的方法，将特征转换为嵌入，并对这些嵌入应用一系列基于注意力的变换。\n17. **[TANGOS](https:\u002F\u002Fopenreview.net\u002Fpdf?id=n6H86gW8u0d)**：一种基于正则化的表格数据方法，利用梯度归因鼓励神经元特化和正交化。\n18. **[SwitchTab](https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.02013)**：一种专为表格数据设计的自监督方法，通过非对称的编码器-解码器框架改进表示学习。遵循原始论文，我们的工具包采用监督学习形式，在每个epoch同时优化重建损失和监督损失。\n19. **[PTaRL](https:\u002F\u002Fopenreview.net\u002Fpdf?id=G32oY4Vnm8)**：一种基于正则化的框架，通过构建并投影到原型空间来提升预测性能。\n20. **[TabPFN](https:\u002F\u002Farxiv.org\u002Fabs\u002F2207.01848)**：一种通用模型，涉及使用预训练的深度神经网络，可直接应用于任何表格任务。\n21. **[HyperFast](https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.14335)**：一种元训练的超网络，可生成特定任务的神经网络，用于即时分类表格数据。\n22. **[TabPTM](https:\u002F\u002Farxiv.org\u002Fabs\u002F2311.00055)**：一种通用的表格数据方法，利用元表示标准化异构数据集，使预训练模型无需额外训练即可泛化到未见数据集。\n23. **[BiSHop](https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.03830)**：一个端到端的深度表格学习框架，利用可调稀疏性的稀疏霍普菲尔德模型，并辅以列级和行级模块。\n24. **[ProtoGate](https:\u002F\u002Farxiv.org\u002Fabs\u002F2306.12330)**：一种基于原型的HDLSS生物医学数据特征选择模型，通过自适应的全局和局部特征选择提升预测精度和可解释性，并借助非参数化的原型机制解决共适应问题。\n25. **[RealMLP](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.04491)**：一种改进的多层感知机（MLP）。\n26. **[MLP_PLR](https:\u002F\u002Farxiv.org\u002Fabs\u002F2203.05556)**：一种改进的多层感知机（MLP），采用周期性激活函数。\n27. **[Excelformer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2301.02819)**：一种用于表格数据预测的深度学习模型，配备半透性注意力模块以应对旋转不变性、定制化数据增强以及注意力前馈网络，使其成为跨不同数据集的可靠解决方案。\n28. **[GRANDE](https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.17130)**：一种模仿树结构的方法，使用端到端梯度下降学习硬性、轴对齐的决策树集成。\n29. **[AMFormer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.02334)**：一种基于标记的方法，通过引入并行加法和乘法注意力机制改进变压器架构用于表格数据，并利用提示标记约束特征交互。\n30. **[Trompt](https:\u002F\u002Farxiv.org\u002Fabs\u002F2305.18446)**：一种基于提示的深度神经网络，用于表格数据，将学习分为内在列特征和样本特定的重要性特征。\n31. **[TabM](https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.24210)**：一种基于MLP及BatchEnsemble变体的模型。\n32. **[TabPFN v2](https:\u002F\u002Fdoi.org\u002F10.1038\u002Fs41586-024-08328-6)**：一种通用模型，涉及使用预训练的深度神经网络，可直接应用于任何表格任务。\n33. **[T2Gformer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.16887)**：一种用于表格学习的Transformer网络，通过关系图引导数据处理，并使用跨层级读出机制获取全局语义用于预测。\n34. **[TabICL](https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.05564)**：一个可比较的表格基础模型，性能与TabPFN v2相当。\n35. **[TabAutoPNPNet](https:\u002F\u002Fwww.mdpi.com\u002F2079-9292\u002F14\u002F6\u002F1165)**：一个基于周期性的表格模型，特别是傅里叶变换和切比雪夫多项式，其性能与FT-Transformer相当或更优。\n36. **[Mitra](https:\u002F\u002Fwww.amazon.science\u002Fblog\u002Fmitra-mixed-synthetic-priors-for-enhancing-tabular-foundation-models)**：一个使用混合合成先验学习的表格基础模型。\n37. **[LimiX](https:\u002F\u002Farxiv.org\u002Fabs\u002F2509.03505)**：一个表格基础模型，利用Transformer支持广泛的任务，从预测到插补和因果推断，所有功能均在一个统一的架构中实现。\n38. **[Real-TabPFN](https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.03971)**：一个增强的表格基础模型，通过对真实世界数据集持续预训练，将TabPFNv2扩展用于分类任务。\n39. **[RFM](https:\u002F\u002Fwww.science.org\u002Fdoi\u002F10.1126\u002Fscience.adi5639)**：一种非深度、无需反向传播的特征学习算法，迭代地将AGOP应用于核机器，以自适应方式学习特定于任务的特征。\n40. **[xRFM](https:\u002F\u002Farxiv.org\u002Fabs\u002F2508.10053)**：一个表格模型，将RFM与自适应树结构相结合，使其能够学习数据子集中的局部特征，并随样本数量呈对数线性增长。\n\n🔧 如果您想查看所有方法的**默认超参数及超参数搜索空间**，请访问：\n👉 [https:\u002F\u002F6sy666.github.io\u002FTALENT-Configs\u002F](https:\u002F\u002F6sy666.github.io\u002FTALENT-Configs\u002F)\n\n\n\n## ☄️ 如何使用 TALENT\n\n### 🕹️ 快速入门\n\n通过 GitHub 安装最新版本：\n\n```bash\n$ pip install git+https:\u002F\u002Fgithub.com\u002FLAMDA-Tabular\u002FTALENT.git@main --upgrade\n```\n\n尝试一个示例 `train_model_deep.py`：\n\n```python\n\nfrom tqdm import tqdm\nfrom TALENT.model.utils import get_deep_args,show_results,tune_hyper_parameters,get_method,set_seeds\nfrom TALENT.model.lib.data import get_dataset\n\nif __name__ == '__main__':\n    loss_list, results_list, time_list = [], [], []\n    args,default_para,opt_space = get_deep_args()\n    train_val_data,test_data,info = get_dataset(args.dataset,args.dataset_path)\n    if args.tune:\n        args = tune_hyper_parameters(args,opt_space,train_val_data,info)\n    for seed in tqdm(range(args.seed_num)):\n        args.seed = seed    # 更新随机种子  \n        set_seeds(args.seed)\n        method = get_method(args.model_type)(args, info['task_type'] == 'regression')\n        time_cost = method.fit(train_val_data, info)    \n        vl, vres, metric_name, predict_logits = method.predict(test_data, info, model_name=args.evaluate_option)\n\t    loss_list.append(vl)\n        results_list.append(vres)\n        time_list.append(time_cost)\n\n    show_results(args,info, metric_name,loss_list,results_list,time_list)\n\n```\n\n\n\n```bash\npython train_model_deep.py --model_type MODEL_NAME\n```\n\n\n\n> 供研究人员参考：\n\n### 🕹️ 克隆\n\n克隆此 GitHub 仓库：\n\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002FLAMDA-Tabular\u002FTALENT\ncd TALENT\u002Ftest\n```\n\n### 🔑 运行实验\n\n1. 编辑 `configs\u002Fdefault\u002F[MODEL_NAME].json` 和 `config\u002Fopt_space\u002F[MODEL_NAME].json`，以配置全局设置和超参数。\n\n2. 运行以下命令：\n    ```bash\n    python train_model_deep.py --model_type MODEL_NAME\n    ```\n    用于深度学习方法；或者：\n    ```bash\n    python train_model_classical.py --model_type MODEL_NAME\n    ```\n    用于传统机器学习方法。\t\n\n### 🛠️ 如何添加新方法\n\n对于像 MLP 类这样只需设计模型的方法，您只需：\n\n- 将模型类添加到 `model\u002Fmodels`。\n- 继承自 `model\u002Fmethods\u002Fbase.py`，并在新类中重写 `construct_model()` 方法。\n- 在 `model\u002Futils.py` 的 `get_method` 函数中添加该方法的名称。\n- 在 `configs\u002Fdefault\u002F[MODEL_NAME].json` 和 `configs\u002Fopt_space\u002F[MODEL_NAME].json` 中添加该新方法的参数设置。\n\n对于需要修改训练流程的其他方法，请基于 `model\u002Fmethods\u002Fbase.py` 部分重写相关函数。具体细节可参考 `model\u002Fmethods\u002F` 中其他方法的实现。\n\n更多详细信息，请参阅我们的[贡献指南](CONTRIBUTING.md)。\n\n### 📦 依赖项\n\n```bash\n   pip install -r requirements.txt\n```\n\n\n如果您想使用 **TabR**，则需要手动安装 faiss，而 faiss 目前仅在 **conda** 环境中可用：\n\n```bash\nconda install faiss-gpu -c pytorch\n```\n\n## 🗂️ 基准数据集\n\n数据集可在 [Google Drive](https:\u002F\u002Fdrive.google.com\u002Fdrive\u002Ffolders\u002F1j1zt3zQIo8dO6vkO-K-WE6pSrl71bf0z?usp=drive_link) 上获取。\n\n### 📂 数据集放置方式\n\n数据集应放置在项目的当前目录下，并与 `args.dataset_path` 指定的文件名相对应。例如，如果项目名为 `LAMDA-TALENT`，则数据应放置在 `LAMDA-TALENT\u002Fargs.dataset_path\u002Fargs.dataset` 中。\n\n每个数据集文件夹 `args.dataset` 包含：\n\n- 数值特征：`N_train\u002Fval\u002Ftest.npy`（如果没有数值特征，可以省略）\n\n- 分类特征：`C_train\u002Fval\u002Ftest.npy`（如果没有分类特征，可以省略）\n\n- 标签：`y_train\u002Fval\u002Ftest.npy`\n\n- `info.json` 文件，其中必须包含以下三项内容（`task_type` 可为“回归”、“多分类”或“二分类”）：\n  \n  ```json\n  {\n    \"task_type\": \"regression\", \n    \"n_num_features\": 10,\n    \"n_cat_features\": 10\n  }\n  ```\n\n## 📝 实验结果\n\n我们在图中基于我们的工具箱，以公平的方式对经典方法和深度表格方法进行了全面评估。考虑了三种表格预测任务，即二分类、多分类和回归，每个子图代表一种不同的任务类型。\n\n我们分别使用`Accuracy`和`RMSE`作为分类任务和回归任务的评价指标。为了校准这些指标，我们选择了平均性能排名来比较所有方法，其中排名越低表示性能越好，遵循[Sheskin (2003)](https:\u002F\u002Fwww.taylorfrancis.com\u002Fbooks\u002Fmono\u002F10.1201\u002F9781420036268\u002Fhandbook-parametric-nonparametric-statistical-procedures-david-sheskin)的方法。效率则通过平均训练时间（以秒为单位）计算，数值越低表示时间效率越高。模型大小通过圆圈的半径直观地表示，便于快速了解模型复杂度与性能之间的权衡。\n\n\u003C!-- TALENT 中提供的经典方法 `SVM` 是一个 `LinearSVM`，以确保更快的训练速度。我们还考虑了 `Dummy` 基线，它会输出主要类别的标签以及分类和回归任务的平均标签。 -->\n\n\u003Cdiv align=\"center\" style=\"text-align:center;\">\n\n  \u003Cdiv style=\"display:inline-block; width:45%; vertical-align:top; margin-bottom:5px;\">\n    \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FLAMDA-Tabular_TALENT_readme_2ebfaaa192f5.png\" alt=\"Binary classification\" style=\"width:45%; border-radius:8px;\" \u002F>\n    \u003Cdiv>(a) 二分类\u003C\u002Fdiv>\n  \u003C\u002Fdiv>\u003Cdiv style=\"display:inline-block; width:45%; vertical-align:top; margin-bottom:5px;\">\n    \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FLAMDA-Tabular_TALENT_readme_da07af822efc.png\" alt=\"Multiclass Classification\" style=\"width:45%; border-radius:8px;\" \u002F>\n    \u003Cdiv>(b) 多分类\u003C\u002Fdiv>\n  \u003C\u002Fdiv>\n\n  \u003Cdiv style=\"display:inline-block; width:45%; vertical-align:top; margin-bottom:5px;\">\n    \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FLAMDA-Tabular_TALENT_readme_97e6e7e0e661.png\" alt=\"Regression\" style=\"width:45%; border-radius:8px;\" \u002F>\n    \u003Cdiv>(c) 回归\u003C\u002Fdiv>\n  \u003C\u002Fdiv>\u003Cdiv style=\"display:inline-block; width:45%; vertical-align:top; margin-bottom:5px;\">\n    \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FLAMDA-Tabular_TALENT_readme_5ce9070ebd8c.png\" alt=\"All tasks\" style=\"width:45%; border-radius:8px;\" \u002F>\n    \u003Cdiv>(d) 所有任务\u003C\u002Fdiv>\n  \u003C\u002Fdiv>\n\n\u003C\u002Fdiv>\n\n\u003C!-- 从对比中可以看出，**CatBoost** 在大多数分类和回归任务中都取得了最佳的平均排名。在所有深度表格方法中，**ModernNCA** 在大多数情况下表现最佳，同时保持了可接受的训练成本。这些结果凸显了 CatBoost 和 ModernNCA 在处理各种表格预测任务时的有效性，使其成为那些寻求高性能和高效率的从业者们的理想选择。\n\n这些可视化图表提供了一种有效且公平的工具，可以快速评估不同任务类型下各类表格方法的优缺点，从而帮助研究人员和从业者根据自身需求做出明智的模型选择决策。 -->\n\n## 👨🏫 致谢\n\n我们感谢以下仓库在我们的工作中提供了有用的组件\u002F函数：\n\n- [Rtdl-revisiting-models](https:\u002F\u002Fgithub.com\u002Fyandex-research\u002Frtdl-revisiting-models)\n- [Rtdl-num-embeddings](https:\u002F\u002Fgithub.com\u002Fyandex-research\u002Frtdl-num-embeddings)\n- [Tabular-dl-tabr](https:\u002F\u002Fgithub.com\u002Fyandex-research\u002Ftabular-dl-tabr)\n- [DANet](https:\u002F\u002Fgithub.com\u002FWhatAShot\u002FDANet)\n- [TabCaps](https:\u002F\u002Fgithub.com\u002FWhatAShot\u002FTabCaps)\n- [DNNR](https:\u002F\u002Fgithub.com\u002Fyounader\u002Fdnnr)\n- [PTaRL](https:\u002F\u002Fgithub.com\u002FHangtingYe\u002FPTaRL)\n- [Saint](https:\u002F\u002Fgithub.com\u002Fsomepago\u002Fsaint)\n- [SwitchTab](https:\u002F\u002Fgithub.com\u002Favivnur\u002FSwitchTab)\n- [TabNet](https:\u002F\u002Fgithub.com\u002Fdreamquark-ai\u002Ftabnet)\n- [TabPFN](https:\u002F\u002Fgithub.com\u002Fautoml\u002FTabPFN)\n- [Tabtransformer-pytorch](https:\u002F\u002Fgithub.com\u002Flucidrains\u002Ftab-transformer-pytorch)\n- [TANGOS](https:\u002F\u002Fgithub.com\u002Falanjeffares\u002FTANGOS)\n- [GrowNet](https:\u002F\u002Fgithub.com\u002Fsbadirli\u002FGrowNet)\n- [HyperFast](https:\u002F\u002Fgithub.com\u002FAI-sandbox\u002FHyperFast)\n- [BiSHop](https:\u002F\u002Fgithub.com\u002FMAGICS-LAB\u002FBiSHop)\n- [ProtoGate](https:\u002F\u002Fgithub.com\u002FSilenceX12138\u002FProtoGate)\n- [Pytabkit](https:\u002F\u002Fgithub.com\u002Fdholzmueller\u002Fpytabkit)\n- [Excelformer](https:\u002F\u002Fgithub.com\u002FWhatAShot\u002FExcelFormer)\n- [GRANDE](https:\u002F\u002Fgithub.com\u002Fs-marton\u002FGRANDE)\n- [AMFormer](https:\u002F\u002Fgithub.com\u002Faigc-apps\u002FAMFormer)\n- [TabM](https:\u002F\u002Fgithub.com\u002Fyandex-research\u002Ftabm)\n- [TabICL](https:\u002F\u002Fgithub.com\u002Fsoda-inria\u002Ftabicl)\n- [TabAutoPNPNet](https:\u002F\u002Fgithub.com\u002Fmatteo-rizzo\u002Fperiodic-tabular-dl)\n- [LimiX](https:\u002F\u002Fgithub.com\u002Flimix-ldm\u002FLimiX)\n- [xRFM](https:\u002F\u002Fgithub.com\u002Fdmbeaglehole\u002FxRFM)\n\n## 🤗 联系方式\n\n如有任何问题，请随时通过提交 issue 提出新功能建议，或联系作者：**刘思洋** ([liusy@lamda.nju.edu.cn](mailto:liusy@lamda.nju.edu.cn))、**蔡浩然** ([caihr@lamda.nju.edu.cn](mailto:caihr@lamda.nju.edu.cn))、**周启乐** ([zhouql@lamda.nju.edu.cn](mailto:zhouql@lamda.nju.edu.cn))、**蒋俊鹏** ([jiangjp@lamda.nju.edu.cn](mailto:jiangjp@lamda.nju.edu.cn))、**殷怀宏** ([yinhh@lamda.nju.edu.cn](mailto:yinhh@lamda.nju.edu.cn))、**周涛** ([zhout@lamda.nju.edu.cn](mailto:zhout@lamda.nju.edu.cn))以及**叶涵嘉** ([yehj@lamda.nju.edu.cn](mailto:yehj@lamda.nju.edu.cn))。祝您使用愉快。\n\n## 🚀 星级历史\n\n[![Star History Chart](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FLAMDA-Tabular_TALENT_readme_5ac421800874.png)](https:\u002F\u002Fstar-history.com\u002F#LAMDA-Tabular\u002FTALENT&Date)\n\n> 感谢 [LAMDA-PILOT](https:\u002F\u002Fgithub.com\u002FLAMDA-CL\u002FLAMDA-PILOT) 和 [LAMDA-ZhiJian](https:\u002F\u002Fgithub.com\u002Fzhangyikaii\u002FLAMDA-ZhiJian) 提供的模板。","# TALENT 快速上手指南\n\nTALENT (Tabular Analytics and Learning Toolbox) 是一个专为表格数据设计的综合机器学习工具箱，集成了先进的深度学习模型、经典算法及高效的超参数调优功能，旨在提升表格数据的模型性能。\n\n## 环境准备\n\n在开始之前，请确保您的系统满足以下要求：\n\n*   **操作系统**: Linux, macOS 或 Windows\n*   **Python 版本**: 3.10 (推荐)\n*   **PyTorch 版本**: 2.0.1 (推荐)\n*   **硬件**: 建议使用支持 CUDA 的 GPU 以加速深度学习模型的训练（可选，CPU 亦可运行部分模型）\n\n**前置依赖检查：**\n请确保已安装 `git` 和 `pip`。\n\n## 安装步骤\n\n### 1. 克隆仓库\n首先，从 GitHub 克隆 TALENT 项目代码：\n\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002Fqile2000\u002FLAMDA-TALENT.git\ncd LAMDA-TALENT\n```\n\n> **国内加速提示**：如果访问 GitHub 速度较慢，可以使用 Gitee 镜像（如果有）或通过配置代理加速下载。\n\n### 2. 创建虚拟环境（推荐）\n为了避免依赖冲突，建议创建独立的 Python 虚拟环境：\n\n```bash\npython -m venv talent_env\nsource talent_env\u002Fbin\u002Factivate  # Linux\u002FmacOS\n# talent_env\\Scripts\\activate   # Windows\n```\n\n### 3. 安装依赖\n安装项目所需的 Python 依赖包。项目通常提供 `requirements.txt` 文件：\n\n```bash\npip install -r requirements.txt\n```\n\n> **国内源加速**：推荐使用清华或阿里镜像源加速安装：\n> ```bash\n> pip install -r requirements.txt -i https:\u002F\u002Fpypi.tuna.tsinghua.edu.cn\u002Fsimple\n> ```\n\n如果需要安装特定版本的 PyTorch（根据官方推荐 2.0.1），请访问 [PyTorch 官网](https:\u002F\u002Fpytorch.org\u002Fget-started\u002Flocally\u002F) 获取对应命令，或手动指定版本安装。\n\n## 基本使用\n\nTALENT 的设计注重易用性，以下是运行一个简单基准测试或训练模型的基本流程。\n\n### 1. 数据集准备\nTALENT 内置了 300+ 个基准数据集。通常无需手动下载，工具箱会自动处理。如需使用自定义数据集，可参考 `data\u002F` 目录下的格式说明。\n\n### 2. 运行示例脚本\n项目通常提供主入口脚本来执行训练或评估。假设我们要在一个默认数据集上运行一个基础模型（如 MLP 或 TabNet）：\n\n```bash\npython main.py --dataset \u003Cdataset_name> --method \u003Cmethod_name>\n```\n\n**具体示例**：\n运行 `MLP` 模型在 `adult` 数据集上进行训练和评估：\n\n```bash\npython main.py --dataset adult --method mlp --config configs\u002Fdefault.yaml\n```\n\n*注：具体的 `\u003Cdataset_name>` 和 `\u003Cmethod_name>` 请参考项目 `methods\u002F` 和 `data\u002F` 目录下的可用列表，或查阅官方文档 [Docs](https:\u002F\u002Flamda-talent.readthedocs.io\u002Fen\u002Flatest)。*\n\n### 3. 查看结果\n运行结束后，模型的性能指标（如 Accuracy, AUC, RMSE 等）通常会输出到终端，并保存至 `results\u002F` 或 `logs\u002F` 目录下，方便后续分析。\n\n### 4. 添加自定义方法（进阶）\nTALENT 支持高度定制化。您可以轻松在 `methods\u002F` 目录下添加新的深度学习架构或经典算法，只需遵循现有的类继承结构即可无缝集成到工具箱的评估流程中。\n\n---\n*更多详细用法、参数配置及最新模型列表（如 ModernNCA, TabPFN v2, RFM 等），请参阅项目的 [官方文档](https:\u002F\u002Flamda-talent.readthedocs.io\u002Fen\u002Flatest) 或 [论文](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.04057)。*","某金融科技公司数据团队正面临构建高精度用户信用评分模型的挑战，需在有限时间内从数百个特征中挖掘最佳预测方案。\n\n### 没有 TALENT 时\n- **模型选型盲目**：团队仅依赖熟悉的 XGBoost 或基础神经网络，难以判断最新的深度学习方法（如 Real-TabPFN）是否更适合当前稀疏的表格数据，容易陷入局部最优。\n- **数据准备繁琐**：面对 300+ 种不同分布的数据集，需手动编写大量代码进行清洗、归一化和编码，预处理流程重复且易出错，消耗了 80% 的开发时间。\n- **基准对比缺失**：缺乏统一的评估框架，无法在相同条件下公平对比 35+ 种深度学习算法与经典方法的性能，导致模型迭代方向模糊。\n- **复现成本高昂**：尝试复现论文中的 SOTA 模型时，常因环境配置复杂或缺少标准超参数设置而失败，研发周期被无限拉长。\n\n### 使用 TALENT 后\n- **智能模型匹配**：直接调用 TALENT 内置的 35+ 种前沿深度方法及 10+ 种经典算法，快速在统一基准下筛选出最适合信用评分任务的模型组合。\n- **自动化预处理**：利用其强大的预处理模块，一键完成多样化的归一化与编码操作，自动适配 300+ 数据集特性，将数据准备时间缩短至原来的 20%。\n- **全方位性能洞察**：基于标准化评测体系，清晰量化各模型在不同指标下的表现，迅速定位到比传统树模型提升 5% AUC 的最佳深度学习架构。\n- **开箱即用体验**：无需纠结环境配置与超参数微调，直接加载预置配置即可复现顶尖论文效果，让团队专注于业务逻辑优化而非底层工程搭建。\n\nTALENT 通过提供标准化的全链路工具箱，将表格数据学习的试错成本降至最低，助力团队在极短时间内交付业界领先的预测模型。","https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FLAMDA-Tabular_TALENT_3ae20edd.png","LAMDA-Tabular","https:\u002F\u002Foss.gittoolsai.com\u002Favatars\u002FLAMDA-Tabular_b75af0b9.png","",null,"https:\u002F\u002Fgithub.com\u002FLAMDA-Tabular",[81,85,89,93],{"name":82,"color":83,"percentage":84},"Python","#3572A5",89.6,{"name":86,"color":87,"percentage":88},"HTML","#e34c26",9.5,{"name":90,"color":91,"percentage":92},"Jupyter Notebook","#DA5B0B",0.8,{"name":94,"color":95,"percentage":96},"Shell","#89e051",0.1,834,55,"2026-04-03T17:08:46","MIT","未说明","未说明（工具包含多种深度学习模型如 TabPFN、Transformer 等，通常建议配备 NVIDIA GPU 以加速训练，但 README 未明确最低显存或 CUDA 版本要求）","未说明（考虑到包含 300 个数据集及大型基准测试，建议 16GB 以上）",{"notes":105,"python":106,"dependencies":107},"README 中明确标注了 PyTorch 2.0.1 和 Python 3.10 的版本徽章。该工具箱集成了 30 多种深度学习架构（包括 TabPFN、TabNet、Transformer 变体等）及 300 个数据集。虽然未明确列出操作系统和 GPU 硬性指标，但鉴于其深度学习特性，建议在支持 CUDA 的 Linux 环境下运行以获得最佳性能。部分预训练模型（如 TabPFN）可能需要较大的内存加载。","3.10",[108,109,110,111],"torch==2.0.1","scikit-learn","pandas","numpy",[54,51,13],[114,115,116,117,118,119,120],"tabular","tabular-data","tabular-methods","tabular-data-benchmark","tabular-data-deep-learning","tabular-data-toolkit","tabular-data-machine-learning","2026-03-27T02:49:30.150509","2026-04-06T05:35:41.421616",[124,129,134,139,144,149],{"id":125,"question_zh":126,"answer_zh":127,"source_url":128},16234,"TALENT 基准测试中分类和回归任务分别使用什么评估指标？","分类任务使用准确率（Accuracy），回归任务使用均方根误差（RMSE）。这与排行榜上的其他结果保持一致。","https:\u002F\u002Fgithub.com\u002FLAMDA-Tabular\u002FTALENT\u002Fissues\u002F74",{"id":130,"question_zh":131,"answer_zh":132,"source_url":133},16235,"为什么 ModernNCA 和其他深度学习模型默认使用双精度（float64）而不是单精度（float32）？","虽然实验表明 ModernNCA 在 float32 和 double precision 之间性能差异不显著，且 float32 效率更高，但团队在其他涉及最近邻搜索的项目（如 TabPTM）中发现精度会影响结果。为了统一性和避免潜在的数值敏感问题，目前所有深度学习方法默认使用双精度。未来版本计划增加让用户选择精度的选项。","https:\u002F\u002Fgithub.com\u002FLAMDA-Tabular\u002FTALENT\u002Fissues\u002F6",{"id":135,"question_zh":136,"answer_zh":137,"source_url":138},16236,"在哪里可以找到每个方法和数据集的最佳超参数配置？","维护者已分享调优结果。注意：如果在 modernNCA 配置中看到 `n_bins` 参数被调优，可以忽略它。这是因为在探索数值编码时为了方便加入的，最终结果基于默认设置（无额外数值编码），该参数不影响最终模型。","https:\u002F\u002Fgithub.com\u002FLAMDA-Tabular\u002FTALENT\u002Fissues\u002F11",{"id":140,"question_zh":141,"answer_zh":142,"source_url":143},16237,"论文《A Closer Look at Deep Learning on Tabular Data》表 3 中的数值代表什么含义？","表 3 中的数值代表“平均排名”（average rank），数值越小表示性能越好。具体计算方式是：针对某种方法（如 MLP）在特定类型数据集上，对其采用的三种不同编码策略（Vanilla, E-Q, E-T）的排名取平均值。","https:\u002F\u002Fgithub.com\u002FLAMDA-Tabular\u002FTALENT\u002Fissues\u002F2",{"id":145,"question_zh":146,"answer_zh":147,"source_url":148},16238,"在使用 TALENT 进行回归任务时，如果模型表现不佳，可能是什么原因导致的？","这可能是由于归一化反转处理的问题。虽然 `LamdaLow.predict(..)` 内部包含了回归的反向归一化逻辑，但在某些自定义封装或基准测试中，如果未正确调用预测接口或未正确处理 `self.method.y_info` 中的归一化参数，会导致指标异常。建议检查是否手动进行了错误的反向归一化操作，或直接使用库提供的预测方法。已有用户创建了复现仓库（LAMDA-TALENT-MNCA-REG-ISSUE）来排查此类嵌套交叉验证下的回归问题。","https:\u002F\u002Fgithub.com\u002FLAMDA-Tabular\u002FTALENT\u002Fissues\u002F64",{"id":150,"question_zh":151,"answer_zh":152,"source_url":128},16239,"是否有包含新加入方法（如 TabPFNv2, TabICL）的最新基准测试结果？","截至该 Issue 讨论时，维护者尚未运行或计划近期运行这些计算密集型新方法（如 TabPFNv2 和 TabICL）的基准测试。用户可以关注项目后续的更新或结果页面（https:\u002F\u002F6sy666.github.io\u002FTALENT-Results\u002F）以获取最新数据。",[]]