Predicting Daily Activities Effectiveness Using Base-level and Meta level Classifiers

Abstract

Collecting and analyzing Activities of Daily Living (ADL) could supplement elder care and long-term care services with very sensitive information about elder people and what they do during the day and what challenges they face. Providing care for elder people based on their ADL could let them live actively, independently and healthy. In this paper, we studied the effectiveness of base learners against ensemble methods for predicting ADL. The selected base learners are Naïve Bayes, Bayesian Network, Sequential Minimal Optimization, Decision Table and J48 while the selected ensemble learners are boosting, bagging, decorate and random forest. The dataset was gathered from a wearable accelerometer attached on the chest. The data used in this study is collected from fifteen participants conducting seven activities namely standing up, working at the computer, going up downstairs, standing, walking, walking and talking with someone and talking while standing, walking and going up downstairs. For base learners, J48 achieved the best results in terms of F-measure, precision and recall. Results also showed that Boosting using decision table as the base classifier achieved the best improvement over base classifier. In addition, Bagging was the only ensemble approach that improved the results using all classifiers as base learners. Moreover, Bagging was able to predict five activities out of seven more efficiently than the other approaches while the rotation forest approach was able to predict the remaining two activities more efficiently than the rest. The results also indicated that all approaches took a reasonable time to build the model except Decorate.