基于时空矩阵分解的空气质量推断

Spatio-temporal Matrix Factorization Based Air Quality Inference

随着城市化进程加快,空气污染形势也愈发严峻,在城市范围内提供细粒度的空气质量时空分布对于人们的户外活动规划和身体健康具有重要意义。然而,稀疏的空气质量站点、不完备的相关特征数据及空气质量随空间和时间的非线性变化使无站点区域的空气质量准确推断面临巨大挑战。通过分析实际空气质量数据集,发现空气质量矩阵的低秩结构,并基于此提出一种基于低秩矩阵分解的方法,通过融合来自低秩结构、空气质量测量值和各类时空特征的信息进行空气质量推断。与现有工作分别处理特征恢复、特征提取和空气质量推断不同,本文方法将3个任务统一到一个模型,通过对不同任务的协同训练与监督提升总体的推断性能。在模型中,构建空间、时间特征矩阵及空气质量矩阵,进行联合低秩矩阵分解,分别得到空间区域的特征表示、不同时刻的特征表示,通过与空气质量矩阵共享空间及时间矩阵因子,将空间和时间特征表达的时空相似性信息迁移到空气质量矩阵缺失值推断以提升其性能。基于北京市的真实空气质量数据集,将所提模型与基线模型进行对比,结果表明所提模型在推断误差、标准差等指标上均优于基线模型,具有较好的FAC2结果,能够在一定程度上揭示影响空气质量变化的主要时空特征。

With rapid urbanization,air pollution has become increasingly severe,making the provision of a spatio-temporal fine-grained air quality distribution essential to support outdoor planning and promote good health.However,the sparseness of air quality stations,the incompleteness of related feature data,and the nonlinear variation of air quality across locations and times pose substantial challenges for accurately inferring air quality in unobserved areas.This study proposes a matrix factorization-based approach to infer air quality by analyzing a real air quality dataset and discovering the low-rank structure of the air quality matrix.This approach fuses knowledge from the low-rank structure,air quality measurements,and various spatio-temporal features.Unlike existing works that address feature recovery,feature extraction,and air quality inference separately,this study unifies these three tasks into a single model.Such integration allows for improved inference performance through the collaborative training and supervision of different tasks.In this model,spatial and temporal feature matrices and the air quality matrix are constructed and collaboratively factorized into spatial and temporal feature representations.By sharing spatio-temporal matrix factors with the air quality matrix,the similarity knowledge of spatial and temporal features is transferred into air quality inference to enhance its performance.The proposed model is evaluated using real data sources obtained in Beijing city.Comparison results with baseline models demonstrate that the proposed model surpasses these models in various metrics,such as inference error and standard deviation,and achieves a better FAC2 result.Additionally,the model effectively reveals the principal spatial and temporal features to a certain extent.