Missing Data Imputation Using Regression Tree Model for Sparse Data Collected via Wide Area Ubiquitous Network

In a ubiquitous/pervasive environment, devices such as sensors and actuators will exist in high density. In this environment, we can acquire a large number of sensor values such as temperature and humidity. We have proposed a ubiquitous data storing architecture called uTupleSpace (uTS), which supports flexible sharing of sensor values with multiple users/software/devices. However, despite a user request, if some values are not stored on the uTS, they should be treated as missing and imputed by estimating such values. We focus on the regression tree imputation method for this problem and show its effectivity for a high-density WAUN environment by regarding multiple sensor values observed at the same time as a spatial dataset. Moreover, we propose a preprocessing method for improving the imputation accuracy in a sparse WAUN environment. We can achieve higher accuracy with our preprocessing method compared to the no-preprocessed and linear interpolation methods. We show the effectivity of our proposed method through experiments.