Imputation method based on recurrent neural networks for the internet of things

Abstract

The Internet of Things (IoT) refers to the new technological paradigm in which sensors and common objects, like household appliances, connect to and interact through the Internet -- This new paradigm, and the use of Artificial Intelligence (AI) and modern data analysis techniques, powers the development of smart products and services; which promise to revolutionize the industry and humans way of living -- Nonetheless, there are plenty of issues that need to be solved in order to have reliable products and services based on the IoT -- Among others, the problem of missing data posses great threats to the applicability of AI and data analysis to IoT applications -- This manuscript shows an analysis of the missing data problem in the context of the IoT, as well as the current imputation methods proposed to solve the problem -- This analysis leads to the conclusion that current solutions are very limited when considering how broad the context of IoT applications may be -- Additionally, this manuscript exposes that there is not a common experimental set up in which the authors have tested their proposed imputation methods; moreover, the experiments found in the literature, lack reproducibility and do not carefully consider how the missing data problem may present in the IoT -- Consequently, the reader will find two proposals in this manuscript: i) an experimental set up to properly test imputation methods in the context of the IoT; and ii) an imputation method that is general enough as to be applied to several IoT scenarios -- The latter is based on Recurrent Neural Networks, a family of supervised learning methods which have excel at exploiting patterns in sequential data and intrinsic association between the variables of data