Missing sensor value estimation method for participatory sensing environment

Abstract

Participatory sensing produces incomplete sensor data. Thus, we have to fill in the gaps of any missing values in the sensor data in order to provide sensor-based services. We propose a method to estimate a missing value of incomplete sensor data. It accurately estimates a missing value by repeating two processes: selecting sensors locally correlated with the sensor that includes the missing value and then updating the training sensor dataset that consist of data from the selected sensors available for multiple regression. This procedure effectively helps to find more suitable neighbor records of a query record from the training sensor dataset and to refine the regression model using the records. It overcomes three problems that other estimation methods have: a decrease in the amount of available training sensor dataset due to missing values, the difficulty in finding similar records of a query due to the “curse of dimensionality,” and the complexity in formalizing the estimation model due to “overfitting.” The main feature of our method is the way it repeatedly prunes inessential sensors while exploiting the anti-monotone property in which the training sensor dataset R' that consist of the sensors V' ⊂ V is larger than the data R that consist of V. Empirical evaluations done using public datasets in which we appended missing values show that our method increases the training sensor dataset for estimation and improves estimation accuracy through repeated sensor selections. Furthermore, we confirmed through a field trial and a life-log enrichment trial, that our method was effective for estimating missing sensor values in a participatory sensing environment.