Increasing Prediction Accuracy for Human Activity Recognition Using Optimized Hyperparameters
Niyati R. Darji, Samuel A. Ajila · 2020
In order to provide context-aware services such as health monitoring and customized energy consumption, smart environment designers need to design robust systems for recognizing the Activities of Daily living (ADL). Once these activities are recognized, the data collected can be used for prediction. For example, energy consumption and other characteristics in the home can be predicted. This is possible if human activity in a smart home can be forecasted. The aim of this research work is to "find the best machine learning algorithm to predict human activities and to use hyperparameters tuning through performance optimization to improve the accuracy of the algorithm" The results of our initial experiments using default hyper-parameters show that Random Forest, compared to four other algorithms (MLP, SVM, Naïve Bayes, and Decision Tree), has the best accuracy at 65.32% for all features and 62.54% for reduced number of features. Tuning four Random Forest hyperparameters through optimization increases the accuracy to 97.9777% for all features and 98.287% for reduced features respectively. Using the optimized Random Forest hyperparameters, 20,000 data points are forecasted with MAE of 0.0098 compared to 0.0445 for Support Vector Machine (SVM).