A classification model for incomplete Internet of Things sensor data in decision-making

Abstract

Data generated by sensors in the Internet of Things (IoT) space experiences data loss. This data loss can be caused by many occurrences including network failure, faulty sensors, duplicates, unreliable sensors, and other factors. This paper attempts to make decisions knowing that there is data loss. We used neural network as a promising approach taking into consideration the nature of data we used. Data used in this paper is fast growing, labelled, and can be lost due to afore mentioned possibilities. Hence, we adopted a supervised learning approach that works better with labelled data. We first theoretically evaluated convolutional neural network (CNN), K-nearest neighbor (KNN), naive Bayes, support vector machine (SVM), logistic regression and long-short term memory (LSTM) as promising and potential classifiers/algorithms for a waste collection use case. There has been a challenge of finding an effective waste collection method. The best performing classification model showed high accuracy, recall, precision, and f1 score when working out lost data and is able tell us what action to take. These two decisions are very important to save cost and/or to protect the environment (citizens) in metropolitan areas. The results indicate that the data loss threshold that can be taken per sample is 40% data loss. The results also indicate which algorithm to use at 10%, 20% and 30% of data loss.