Analyzing traffic related geospatial data often lacks in priori knowledge and encounters parameter setting problems due to the dynamic characteristics of city traffic. In this paper, we propose a pervasive, scalable framework for city traffic related geospatial data analysis based on a stacked generalization. Firstly we analyze the optimal linear combination based on stepwise iteration, and also prove its theoretical validity via error-ambiguity decomposition. Secondly we integrate six classical approaches into this framework, including linear least squares regression, autoregressive moving average, historical mean, artificial neural network, radical basis function neural network, support vector machine, and conduct experiments with a real city traffic detecting dataset. We further compare the proposed framework with other four linear combination models. It suggests that the proposed framework behaves more robust than other models both in variance and bias, showing a promising direction for city traffic related geospatial data analysis.
[1]
Zhi-Hua Zhou,et al.
Ensemble Methods: Foundations and Algorithms
,
2012
.
[2]
Manfred K. Warmuth,et al.
On Weak Learning
,
1995,
J. Comput. Syst. Sci..
[3]
Robert L. Winkler,et al.
Simple robust averages of forecasts: Some empirical results
,
2008
.
[4]
T. Evgeniou,et al.
To combine or not to combine: selecting among forecasts and their combinations
,
2005
.
[5]
David H. Wolpert,et al.
Stacked generalization
,
1992,
Neural Networks.
[6]
Hengcai Zhang,et al.
Intersection delay estimation from floating car data via principal curves: a case study on Beijing’s road network
,
2013,
Frontiers of Earth Science.