An evaluation of dynamic selection robustness in noisy environments for activity recognition in smart homes

Abstract

Smart homes can be defined as environments monitored by sensors that capture information executed in it. These sensors are responsible for measure the temperature of a room, the number of times a switch has been turned on, and so on. However, the data obtained in these scenarios may vary during or after the capture process. These variations are defined as noise and affect the interpretation of the data. Given the information obtained from the environment, machine learning techniques can use this knowledge to identify the activities and predict future ones. This area of learning is named Activity Recognition. In recent studies, the Random Forest presented consistent results in Activity Recognition problems in noisy-free environments. To identify which techniques can be used in noisy scenarios, this dissertation evaluated the use of Multiple Classifier Systems in comparison to Random Forest. The proposal is to investigate how these techniques perform on real-world data sets for activity recognition considering six noise levels: 0% to 50%, which refers to a randomly changing in the label activities. Experimental results have shown that the Dynamic Selection techniques are adequate to handle noisy environments presenting stable results as the noise level increases. The performance of OLA and MCB was significantly better than Random Forest even with the 50% noise level.