SCFNet:一种面向时空预测的外部空间特征融合框架  

SCFNet:Fusion Framework of External Spatial Features for Spatio-temporal Prediction

作　　者：刘腾飞 陈李越 房江祎 王乐业 LIU Tengfei;CHEN Liyue;FANG Jiangyi;WANG Leye(Key Laboratory of High Confidence Software Technologies(Peking University),Ministry of Education,Beijing 100871,China;School of Computer Science,Peking University,Beijing 100871,China)

机构地区：[1]高可信软件技术教育部重点实验室(北京大学),北京100871 [2]北京大学计算机学院,北京100871

出　　处：《计算机科学》2025年第4期110-118,共9页Computer Science

摘　　要：道路信息与当前道路的流量模式息息相关,丰富的POI(Point of Interest)语义可以揭示一个地区的属性,人口数据可以揭示一个地区的人口流量趋势。在时空预测中考虑以上外部空间特征对流量带来的影响,可以帮助模型完成更精准的预测。现有的外部空间建模方法通常针对输入的外部空间特征,经过神经网络映射学得空间相关语义表示,再与最终的时空流量表示融合。然而,由于流量表示和空间特征之间具有异构性,已有的外部空间特征建模方法往往扩展性不高,只能针对特定外部空间特征或特定时空模型。为解决以上问题,提出了一种针对外部空间特征的通用建模框架SCFNet(Spatial Context Fusion Network for Traffic Forecasting)。具体而言,引入基于信息交互的注意力机制,在时空表示和外部空间特征之间计算注意力分数,从而实现外部空间特征和时空表示的高效融合;同时,设计了一种时间向量动态编码方式,以生成动态的空间特征语义。SCFNet采用模块化设计,能够与各类最新的时空流量预测网络结合。SCFNet支持区域人口数据、道路信息、POI等不同空间静态特征的混合输入。在3个真实交通数据集上进行了实验,实验结果表明,SCFNet可显著提高各类最新时空预测方法(如MTGNN,ASTGCN,GraphWaveNet)的预测精度。Road information is closely related to the current traffic pattern of roads.Rich POI semantics can reveal the attributes of an area.Demographic data reveals the trend of population flow in an area.Considering the influence brought by the above external spatial features on the flow in spatio-temporal prediction can help the model accomplish more accurate prediction.Existing external spatial modeling methods usually focus on the input external spatial features,learn spatially relevant semantic representations through neural network mapping,and then fuse them with the final spatiotemporal flow representations.However,due to the heterogeneity between flow representations and spatial features,the existing external spatial feature modeling methods are often not highly scalable and can only target specific external spatial features or specific spatio-temporal models.To overcome the above problems,we propose a spatial context fusion network for traffic forecasting(SCFNet).Specifically,we introduce an attention mechanism based on information interaction to compute attention scores between spatio-temporal representations and external spatial features to achieve an efficient fusion of external spatial features and spatio-temporal representations,and we design a dynamic encoding method of time vectors to generate dynamic spatial feature semantics.SCFNet supports a mixture of different spatial static features such as regional demographic data,road information,and POI inputs.We conduct experiments on three real traffic datasets and demonstrated that SCFNet can significantly improve the prediction accuracy of various state-of-the-art spatiotemporal prediction methods such as(MTGNN,ASTGCN,and GraphWaveNet).

关 键 词：时空预测 外部特征建模 兴趣点 道路特征 人口数量 注意力机制 时变语义 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]