Evaluating Performance of SMOTE and ADASYN to Classify Falls and Activities of Daily Living

Class Real-life applications generate skewed data that leads to degraded performance of machine learning models. These methods are biased towards the majority class and cause overfitting problems. Oversampling techniques are used by researchers to balance the dataset by increasing minority class samples which increases the performance of the classifier. Detecting falls and activities of daily living accurately improves the quality of life, particularly for elderly people. In this study, real-time data pertaining to falls and activities are collected using a working prototype. Experiments are performed to balance the data using SMOTE and ADASYN oversampling techniques. Three machine learning models, KNN, RF and CART, with and without parameter tuning, are used for the classification of different types of activities and falls. The results prove that ADASYN and SMOTE improve the accuracy of classification and ADASYN performs better than SMOTE.