Extension of Geographically Weighted Regression for Big Spatial-Temporal Data

• Activity Number: 76
• Type: Contributed
• Date/Time: Sunday, July 29, 2012 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Shu-Ngai Yeung*+ and S. Tom Au and Guangqin Ma
• Companies: AT&T Labs and AT&T Labs and AT&T Labs
• Keywords: Geographically Weighted Regression ; Spatial-Temporal Data ; Data Mining ; Business Applications ; Generalized Least Square

Abstract

As big spatial-temporal dataset becoming common in various domain areas, modeling such dataset is essential for many business applications. While the non-parametric nature of Geographically Weighted Regression (GWR) permits modeling and computational flexibility, GWR is mostly focused on cross-sectional spatial modeling. This paper proposes a general framework for regression modeling on spatial-temporal data by extending the GWR to additional dimensions including the temporal one and dimensions of other attributes. The methodology treats GWR as generalized least squared estimation (GLS) while recursively use the whole dataset when modeling a specific location. Under such view, the variance structure can be generalized to model various attributes. For illustration, this method is applied on real wireless network data by comparing different forms of spatial-temporal weighting functions. The results show that the spatial-temporal weighted regression not only improves the prediction accuracy but also provides a basis for spatial clustering of data usage growth areas.