Exploring the value of the Bregman Block Average Co-clustering algorithm for missing value imputation in geo-referenced time series

Abstract

Introduction Missing values frequently introduce loss of information in spatial analysis. A common approach to manage missing values is to impute missing values. This is often done by using spatial interpolation models, and more recently machine learning methods. The Bregman Block Average Co-clustering algorithm with I-Divergence (BBAC-I) has recently been applied to explore spatial patterns. Among other things, the original authors of this algorithm used it for missing value imputation. This thesis explored the value of the BBAC-I algorithm in missing value imputation of Geo-referenced time series.

Methods This model comparison study compared the imputation value of a selection of machine learning and spatial interpolation models to the BBAC-I models on four data sets with distinctly different spatial characteristics. Three objectives were set to explore the BBAC-I algorithm in this context: (1) Compare the prediction accuracy, (2) compare the computational run time, (3) analyse the spatial properties of the prediction residuals.

Results and Conclusion BBAC-I produced less accurate results than the selection of Machine learning models, but produced more accurate than spatial interpolation methods. The BBAC-I run time was faster than any other model, especially for larger data sets. However, it did consistently produce positively spatially correlated residuals. The value of BBAC-I for missing value imputation lies in a limited selection of data sets that are very large, and for which limiting computational requirements is more important than accuracy. Future research should continue to address the value of recently developed non spatial models in the spatial domain.