TOTRCD

Hugging Face2026-07-25 更新2026-07-27 收录

表格回归评估

概念漂移适应

数据链接：

https://huggingface.co/datasets/yungak/TOTRCD 数据链接链接失效反馈

官方服务：

资源简介：

TOTRCD（Temporally-Ordered Tabular Regression Benchmark Suite with Concept Drift）是一个专为表格回归任务设计的基准测试数据集套件，其核心特点是所有数据都包含时间顺序（通过时间列体现）并呈现出概念漂移现象。该套件旨在为研究时间相关的回归模型以及概念漂移的检测与适应提供标准化的评估资源。数据内容方面，TOTRCD整合了来自真实世界的多个公开表格数据集，涵盖空气质量、航班、能源预测、交通流量、房价、天气、医疗监测等多个领域。所有数据集均经过统一格式化处理，字段名称使用标准化前缀进行标识：`Meta::`表示元数据（如标识符、时间戳等，不作为模型输入）；`Covariate::Static::`表示静态协变量（时间不变的特征）；`Covariate::Temporal::`表示从时间信息派生出的周期性编码特征；`Target::`表示需要预测的目标变量。对于分类变量，采用独热编码，并在列名中使用`::dummy::`中缀进行标记。数据规模上，该套件包含19个子数据集，观测数量从数千到数亿不等（例如，Airlines数据集约有2.26亿条观测，而Coffee Distribution数据集约有6000条观测），特征数量在8到128个之间（经过独热编码后少于1000个）。部分数据集包含多个预测任务（即多个目标变量）。套件中的每个数据集都经过了概念漂移检测流程的验证，确认存在漂移现象。该数据集适用于表格回归任务的研究与评估，特别适合用于探索和测试模型在存在时间依赖性和概念漂移的真实场景下的性能。其严格的纳入标准（包括必须是真实数据、包含时间列、观测数大于5000、特征数有限且存在已验证的概念漂移）保证了数据集的实用性和针对性。

TOTRCD (Temporally-Ordered Tabular Regression Benchmark Suite with Concept Drift) is a benchmark dataset suite specifically designed for tabular regression tasks. Its core characteristic is that all datasets contain temporal order (reflected by the time column) and exhibit concept drift. This suite aims to provide standardized evaluation resources for researching time-related regression models as well as the detection and adaptation of concept drift. In terms of data content, TOTRCD integrates multiple real-world public tabular datasets, covering domains such as air quality, flight, energy forecasting, traffic flow, housing prices, weather, and medical monitoring. All datasets have undergone unified formatting processing, with field names identified using standardized prefixes: `Meta::` denotes metadata (such as identifiers, timestamps, etc., which are not used as model inputs); `Covariate::Static::` denotes static covariates (time-invariant features); `Covariate::Temporal::` denotes periodic encoded features derived from temporal information; `Target::` denotes the target variable to be predicted. For categorical variables, one-hot encoding is adopted, and the `::dummy::` infix is used for marking in column names. In terms of data scale, this suite includes 19 subdatasets, with the number of observations ranging from thousands to hundreds of millions. For example, the Airlines dataset contains approximately 226 million observations, while the Coffee Distribution dataset has approximately 6,000 observations. The number of features ranges from 8 to 128 (fewer than 1,000 after one-hot encoding). Some datasets contain multiple prediction tasks (i.e., multiple target variables). Each dataset in the suite has been verified through a concept drift detection workflow, confirming the presence of concept drift. This dataset is applicable to the research and evaluation of tabular regression tasks, and is particularly suitable for exploring and testing the performance of models in real-world scenarios with temporal dependencies and concept drift. Its strict inclusion criteria—including being real-world data, containing a time column, having more than 5,000 observations, limited feature count, and verified concept drift—ensure the practicality and targeted utility of the dataset.

创建时间：

2026-07-22

原始信息汇总

TOTRCD: 时序表格回归基准套件（含概念漂移）

数据集概述

TOTRCD 是一个包含时间顺序（通过时间列表示）且表现出某种形式概念漂移的表格回归数据集集合。

数据字段结构

数据集列名遵循统一的命名格式，使用前缀标记不同类型的数据：

Meta::：标记标识符、时间戳或排序键，不作为模型输入。例如：Meta::DateTime
Covariate::Static::：标记时间不变的特征输入。例如：Covariate::Static::Distance
Covariate::Temporal::：标记从日期时间得到的周期性编码特征。例如：Covariate::Temporal::Hour (sin)
Target::：标记待预测的目标变量。例如：Target::traffic_volume
::dummy::：插入在分类变量原名和类别值之间，标记独热编码列。例如：Covariate::Static::weather::dummy::Rain

数据集规模

数据集	样本数	特征数	Mahalanobis	ASO	ADWIN
Air Quality	8,991	7,344	7,393	7,396	12
Airlines	226,082,661	8	✅	❌	✅
Appliances Energy Prediction	19,735	29	✅	✅	✅
Beijing PM2.5	41,757	16	✅	✅	✅
Bike Sharing (Washington DC)	17,379	13	✅	✅	✅
CMAPSS	53,759	61,249	25	✅	✅
Coffee Distribution	6,016	103	✅	✅	✅
Gas Turbine Emission	36,733	36,733	9	✅	✅
Marine Cargo Vessel Power Consumption	567,442	10	✅	✅	✅
Metro Interstate Traffic Volume	47,942	22	✅	✅	✅
Miami Housing 2016	13,932	12	✅	✅	✅
NOAA Weather	19,515	12	✅	✅	✅
Parking Birmingham	35,705	35	✅	✅	✅
Parkinsons Telemonitoring	5,875	5,875	19	✅	✅
Seoul Bike Sharing Demand	8,465	18	✅	✅	✅
Shifts Weather	3,544,637	128	✅	❌	✅
Steel Industry Energy Consumption	35,040	15	✅	✅	✅
Temperature Forecast	7,588	7,588	46	✅	✅
Tetouan City Power Consumption	52,416	52,416	52,416	11	✅

注：竖线 | 分隔与同一数据集关联的不同任务（即不同目标标签）。

数据集创建标准

入选该套件的数据集必须满足以下所有条件：

表格回归任务
真实世界数据集，非合成数据
包含一个引入自然顺序的时间列
在允许重新分发的许可证下公开可用
超过5,000个观测值
独热编码后特征少于1,000个
存在概念漂移证据，并经漂移检测流程确认

许可信息

各数据集的许可证信息如下：

数据集	许可证
Air Quality	CC BY 4.0
Airlines	CC0
Appliances Energy Prediction	CC BY 4.0
Beijing PM2.5	CC BY 4.0
Bike Sharing (Washington DC)	CC BY 4.0
CMAPSS	Public Domain
Coffee Distribution	Public Domain
Gas Turbine Emission	CC BY 4.0
Marine Cargo Vessel Power Consumption	CC BY-NC-SA 4.0
Metro Interstate Traffic Volume	CC BY 4.0
Miami Housing 2016	CC BY-NC-SA 4.0
NOAA Weather	CC0
Parking Birmingham	CC BY 4.0
Parkinsons Telemonitoring	CC BY 4.0
Seoul Bike Sharing Demand	CC BY 4.0
Shifts Weather	CC BY-NC-SA 4.0
Steel Industry Energy Consumption	CC BY 4.0
Temperature Forecast	CC BY 4.0
Tetouan City Power Consumption	CC BY 4.0

引用信息

各数据集的原始论文引用如下：

Air Quality

@article{Vito2008OnFC, title={On field calibration of an electronic nose for benzene estimation in an urban pollution monitoring scenario}, author={Saverio De Vito and Ettore Massera and Marco Piga and Luca Martinotto and Girolamo Di Francia}, journal={Sensors and Actuators B-chemical}, year={2008}, volume={129}, pages={750-757}, url={https://api.semanticscholar.org/CorpusID:94886265} }

Appliances Energy Prediction

@article{Candanedo2017DataDP, title={Data driven prediction models of energy use of appliances in a low-energy house}, author={Luis M. Ibarra Candanedo and Veronique Feldheim and Dominique Deramaix}, journal={Energy and Buildings}, year={2017}, volume={140}, pages={81-97}, url={https://api.semanticscholar.org/CorpusID:63814994} }

Beijing PM2.5

@article{Liang2015AssessingBP, title={Assessing Beijings PM2.5 pollution: severity, weather impact, APEC and winter heating}, author={Xuan Liang and Tao Zou and Bin Guo and Shuo Li and Haozhe Zhang and Shuyi Zhang and Hui Huang and Song Xi Chen}, journal={Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences}, year={2015}, volume={471}, url={https://api.semanticscholar.org/CorpusID:130615236} }

Bike Sharing (Washington DC)

@article{FanaeeT2013EventLC, title={Event labeling combining ensemble detectors and background knowledge}, author={Hadi Fanaee-T and Jo{~a}o Gama}, journal={Progress in Artificial Intelligence}, year={2013}, volume={2}, pages={113 - 127}, url={https://api.semanticscholar.org/CorpusID:256282956} }

Gas Turbine Emission

@article{Kaya2019PredictingCA, title={Predicting CO and NOx emissions from gas turbines: Novel data and a benchmark PEMS}, author={Kaya, Heysem and T{"u}fekci, P{i}nar and Uzun, Erdin{c{c}}}, journal={Turkish Journal of Electrical Engineering and Computer Sciences}, year={2019}, volume={27}, number={6}, pages={4783--4796}, doi={10.3906/elk-1807-87} }

Marine Cargo Vessel Power Consumption

@dataset{malinin_2022_7684813, title={Shifts Marine Cargo Vessel Power Consumption Prediction Dataset}, author={Malinin, Andrey and Athanasopoulos, Andreas and Barakovic, Muhamed and Bach Cuadra, Meritxell and Gales, Mark and Granziera, Cristina and Graziani, Mara and Kartashev, Nikolay and Kyriakopoulos, Konstantinos and Lu, Po-Jui and Molchanova, Nataliia and Nikitakis, Antonis and Raina, Vatsal and La Rosa, Francesco and Sivena, Eli and Tsarsitalidis, Vasileios and Tsompopoulou, Efi and Volf, Elena}, publisher={Zenodo}, month={sep}, year={2022}, version={2.0}, doi={10.5281/zenodo.7684813}, url={https://doi.org/10.5281/zenodo.7684813} }

Metro Interstate Traffic Volume

@misc{metro_interstate_traffic_volume_492, author = {Hogue, John}, title = {{Metro Interstate Traffic Volume}}, year = {2019}, howpublished = {UCI Machine Learning Repository}, note = {{DOI}: https://doi.org/10.24432/C5X60B} }

Miami Housing 2016

@techreport{Mayer2021StructuredAR, title={Structured Additive Regression and Tree Boosting}, author={Mayer, Michael and Bourassa, Steven C. and Hoesli, Martin and Scognamiglio, Donato}, institution={Swiss Finance Institute}, type={Swiss Finance Institute Research Paper}, number={21-83}, year={2021}, doi={10.2139/ssrn.3924412}, url={https://ssrn.com/abstract=3924412} }

Parking Birmingham

@inproceedings{Stolfi2017PredictingCP, title={Predicting Car Park Occupancy Rates in Smart Cities}, author={Stolfi, Daniel H. and Alba, Enrique and Yao, Xin}, booktitle={Smart Cities: Second International Conference, Smart-CT 2017}, address={M{a}laga, Spain}, pages={107--117}, year={2017}, doi={10.1007/978-3-319-59513-9_11} }

Parkinsons Telemonitoring

@article{Tsanas2009AccurateTO, title={Accurate Telemonitoring of Parkinsons Disease Progression by Noninvasive Speech Tests}, author={Athanasios Tsanas and Max A. Little and Patrick E. McSharry and Lorraine O. Ramig}, journal={IEEE Transactions on Biomedical Engineering}, year={2009}, volume={57}, pages={884-893}, url={https://api.semanticscholar.org/CorpusID:7382779} }

Seoul Bike Sharing Demand

@article{Sathishkumar2020UsingDM, title={Using data mining techniques for bike sharing demand prediction in metropolitan city}, author={Sathishkumar, V E and Park, Jangwoo and Cho, Yongyun}, journal={Computer Communications}, year={2020}, volume={153}, pages={353--366}, doi={10.1016/j.comcom.2020.02.007} }

@article{Sathishkumar2020ARB, title={A rule-based model for Seoul Bike sharing demand prediction using weather data}, author={Sathishkumar, V E and Cho, Yongyun}, journal={European Journal of Remote Sensing}, year={2020}, volume={53}, number={sup1}, pages={166--183}, doi={10.1080/22797254.2020.1725789} }

Shifts Weather

@inproceedings{ malinin2021shifts, title={Shifts: A Dataset of Real Distributional Shift Across Multiple Large-Scale Tasks}, author={Andrey Malinin and Neil Band and Yarin Gal and Mark Gales and Alexander Ganshin and German Chesnokov and Alexey Noskov and Andrey Ploskonosov and Liudmila Prokhorenkova and Ivan Provilkov and Vatsal Raina and Vyas Raina and Denis Roginskiy and Mariya Shmatova and Panagiotis Tigas and Boris Yangel}, booktitle={Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2)}, year={2021}, url={https://openreview.net/forum?id=qM45LHaWM6E} }

Steel Industry Energy Consumption

@article{VE2020EfficientEC, title={Efficient energy consumption prediction model for a data analytic-enabled industry building in a smart city}, author={Sathishkumar V E and Changsun Shin and Yongyun Cho}, journal={Building Research & Information}, year={2020}, volume={49}, pages={127 - 143}, url={https://api.semanticscholar.org/CorpusID:224916577} }

Temperature Forecast

@misc{bias_correction_of_numerical_prediction_model_temperature_forecast_514, title = {{Bias correction of numerical prediction model temperature forecast}}, year = {2020}, howpublished = {UCI Machine Learning Repository}, note = {{DOI}: https://doi.org/10.24432/C59K76} }

Tetouan City Power Consumption

@article{Salam2018ComparisonOM, title={Comparison of Machine Learning Algorithms for the Power Consumption Prediction : - Case Study of Tetouan city –}, author={Abdul Rahim Salam and Abdelaaziz El Hibaoui}, journal={2018 6th International Renewable and Sustainable Energy Conference (IRSEC)}, year={2018}, pages={1-5}, url={https://api.semanticscholar.org/CorpusID:145050098} }

搜集汇总

数据集介绍

构建方式

在时序回归与概念漂移研究领域，TOTRCD数据集应运而生。其构建遵循一套严谨的入选准则：首先，所有数据必须来源于真实世界的表格型回归任务，而非合成数据；其次，数据需包含一个可赋予自然时序顺序的时间列；同时，为确保统计可靠性，观测样本须超过5000条，且独热编码后的特征数少于1000个。更重要的是，每个数据集必须通过漂移检测管道的验证，确认存在概念漂移现象。基于这些标准，从公开许可的数据库中精选出涵盖空气质量、能源消耗、交通流量等多元领域的19个数据集，构成了该基准套件。

特点

该基准套件的核心特色在于其统一的结构化格式与明确的时序特征。所有数据集均采用一致的列命名前缀体系：‘Meta::’标识不参与建模的元数据（如时间戳），‘Covariate::Static::’与‘Covariate::Temporal::’分别区分静态和时序输入特征，‘Target::’则明确标注预测目标。此外，分类变量通过‘::dummy::’中缀进行独热编码标识。这一精心设计的命名规范极大提升了数据集的可用性与跨数据集分析的互操作性。数据集规模跨度极大，从数千条观测的小型样本到超过2.26亿条记录的大型数据集，为不同规模的实验提供了丰富的选择。

使用方法

研究人员可直接从HuggingFace平台加载TOTRCD数据集，每条记录都保留了原始的时间顺序，便于进行时序预测或概念漂移分析。使用时，用户应关注以‘Meta::’为前缀的列作为排序或分组键，以‘Covariate::’开头的列作为模型输入特征，而‘Target::’列则为回归目标。由于数据集中已包含经漂移检测算法（如ADWIN）验证的标签，研究者可以便捷地设计实验来评估模型在数据分布随时间变化时的鲁棒性。每个子数据集均附有引用信息，鼓励用户在发表成果时引用原始研究以尊重学术贡献。

背景与挑战

背景概述

在时序预测与回归分析的交叉领域，概念漂移（Concept Drift）的存在常使传统静态模型的泛化能力大打折扣。为此，TOTRCD（Temporally-Ordered Tabular Regression Benchmark Suite with Concept Drift）数据集应运而生，由研究团队于近期构建并发布在HuggingFace平台上。该数据集精心整合了19个真实世界的表格回归任务，涵盖空气质量、能源消耗、交通流量、医疗监测等多元场景，每个子集均通过严格筛选确保其包含时间列并表现出可验证的概念漂移现象。其核心研究问题在于为机器学习社区提供一个标准化基准，以系统评估和推动在时间顺序数据中应对分布偏移的算法发展，对稳健预测与自适应学习领域具有重要的示范与引导作用。

当前挑战

该数据集所攻克的领域挑战在于真实环境中数据分布的非平稳性，即概念漂移，它使得基于静态假设的模型在长期预测中性能骤降，这在气象预报、工业监控等持续运行系统的应用中尤为棘手。构建过程中同样面临重重困难：首先，需从海量公开数据中筛选符合表格回归、含时间戳且非合成的子集，样本量跨越五千至两亿余条，规模和维度差异悬殊；其次，必须借助马氏距离、自适应滑动窗口等检测手段逐一验证每个数据源的概念漂移证据，并统一处理周期特征编码与类别变量独热编码等异构格式，最终在遵守各自开源许可协议前提下完成合规的集中发布。

常用场景

经典使用场景

在时间序列预测与概念漂移研究交汇的领域中，TOTRCD数据集作为一项精心构建的基准套件，专门服务于带上时间顺序的表格回归任务。其经典使用场景体现在利用时间列驱动模型对连续目标变量进行预测，并同时验证模型在分布漂移下的鲁棒性。研究者借助套件内集成的马氏距离、自适应滑动窗口等漂移检测工具，能够系统性地评估传统回归模型与在线学习算法在动态环境中的表现差异。

衍生相关工作

围绕TOTRCD数据集，一系列衍生工作已在自适应回归与漂移鲁棒建模领域蓬勃展开。研究者们以其为基准，系统地对比了滑动窗口更新、增量梯度下降、集成学习漂移适应等策略在18类多元回归任务上的效能。此外，该套件还催生了针对时间感知特征工程与周期性编码剥离效果的专项分析，以及基于ADWIN在线检测的模型重训练触发机制优化工作，推动了非平稳环境建模方法论的系统性进步。

数据集最近研究