Accuracy-Aware Data Reduction for Internet of Things

Abstract

The Internet of Things (IoT) generates an enormous amount of data. Usually, either in real-time or after their collection, these data are passed to already trained machine learning and deep learning models for making predictions. Sometimes these data are used for retraining models. While efforts for data reduction have been made, little attention has been given to the effects of data reduction on machine learning models in the IoT context. Furthermore, the works that have explored how accurately data is reduced have no mechanism to use the result as feedback for future data reduction. In this paper, we propose a new reduction technique that depends on feedback from machine learning and deep learning models. As a result, data is as much reduced as required to achieve the desired results for the models. When machine learning model feedback reports low performance, the reduction algorithm increases the amount of data sent. Similarly, when the feedback indicates that the data being sent is sufficient to achieve the desired level of performance, the proposed reduction algorithm attempts to reduce the amount of data further to explore the opportunity that a similar level of performance might be achieved with a comparatively smaller amount of data. Experiments on NASA's turbofan engine data revealed that the algorithm's reduction rate can be integrated with feedback from the machine learning model without significantly affecting the performance of the model.