中图分类号：TP391.7 文献标识码：A DOI：10.7535/hbkd.2025yx03002

Adaptive multi classifier traffic data interpolation model based on information entropy

ZHANG Yunkai1²， ^1，2 ， 1，²， （1.Hebei University Road Traffic Perception and Intellgent Application Technology Researchand Develop Center. Shijiazhuang，Hebei O50035，China； 2.Department of Electrical and Information Engineering，Hebei Jiaotong Vocational and Technical College， Shijiazhuang，Hebei O50035，China； 3.School of Artificial Intelligence and Data Science，Hebei University of Technology， Tianjin 300131，China）

Abstract：Toaddress theisse that single traficdata misingvalue imputation modelscannotcomprehensively handle the multi-sourceheterogeneityandcomplexdata volumeof trafficdata，amulti-clasifierimputation modelbasedonadaptive weighting determined by informationentropywas proposed.First，information entropyrepresenting \"disorder degree\"was introduced toevaluate predictionqualityanddeterminemulti-clasifierweights.Second，adynamicadaptive weightingmethod was designedtoresolve theproblemof differentclassfiers being suitableforvarious samples caused bydeviceheterogeneity. Finall，validationwasconductedonbothpublicandself-collcteddatasets.Theresultsdemonstratethattheproposedmodel achieves significant improvementindetection performancecompared withotherimputation models.Italsoatains highacuracy in experiments on the public Interstate Highway Trafic Flow Dataset，with an F1 of O.778 and a 10% improvement in RMSE，exhibiting strong generalizability.By enabling weights toadaptively evolve withdata streams basedon information entropydetermination，thealgorithmachieves faster detectionspeedand higher accuracy，providing technical references for the establishment of missing value imputation models in traffic data cleaning.

Keywords： data processing；traffc data cleaning;； mising value prediction; information entropy;adaptive weight

随着智慧高速公路的不断发展，高速公路部署了众多终端监测设备来采集种类繁多的数据，如道路数据、车辆数据、气象数据等。这些海量数据存贮在不同的信息系统中，数据之间关系复杂，规模巨大，数据维度高[1]。由于离散的数据采集异常以及通信时数据压缩导致的缺失、设备故障等原因，造成高速公路采集的多源数据中往往会出现数据错误[2]、数据缺失、数据冗余[3]等情况。

现有交通数据清洗方法可分为离线填补与在线处理2类[4]。传统的离线填补方法可进一步划分为基于近邻的填补[5-8]和基于回归的填补[9-13]。基于近邻的KNN、Hot-deck 等算法通过相似性匹配填补缺失值，但在稀疏数据场景下可能产生较大误差。GBKII通过灰色关联度查找缺失值的近邻进行填补;kNNE集成多个特征的相似度来确定近邻填补缺失值;ILLS 在近邻上建立模型以估计缺失值。基于回归的线性模型和核函数模型[14-15]试图挖掘全局相关性，但难以应对数据异构性问题。混合模型如LOESS[16]、ERACER[17]和IIM^[10] 等通过局部回归或多模型融合提升性能，其中IIM在多源交通数据中表现最佳。深度学习技术的引人拓展了数据清洗的边界，张伟光等[18]提出的生成对抗网络、陈俊扬等[19]的图神经网络与强化学习模型及李国等[20]的深度多任务学习框架在特定场景下取得了进展。

然而，现有算法存在两大局限：一是多数方法针对静态数据集，难以适应实时数据流；二是单一模型难以处理交通数据的多源异构性和时空关联性。……