Simplicity is best: addressing the computational cost of machine learning classifiers in constrained edge devices

摘要

Implementing data processing toward the edge of the Internet of Things (IoT) is an important requirement in order to produce real-time feedback according to some decision-based approach. The authors aim to experimentally prove that "simplicity is best." Their experiment-based paper considers three platforms (Notebook i7-7500U, Raspberry PI 3 model B+, Raspberry Zero W) and seven machine learning algorithms for classification (logistic regression, random forest, k-nearest neighbors, naive Bayes, linear support vector machine, multilayer perceptron, and decision trees). They use the Scikit-learn Python library and a dataset for ADL recognition with wrist-worn accelerometer measurements with 839 instances and 14 classes (available on the UCI repository). All results were obtained by training with over 80 percent of data along a five-step cross-validation process. The applied methodology deals with the signals generated by a three-axial inertial sensor (X,Y,Z). After preprocessing, for every component of the signal, five segments are considered. Nine features are computed for each segment, as well for the entire sequence: min, max, average, median, kurtosis, skewness, standard deviation, variance, and mean absolute deviation. Dimension reduction processes, based on chi scores, are used to select first the top ten features and then the top three features.