Combining Clustering Undersample and Ensemble Learning for Wearable Fall Detection

Accidental falls often cause serious harm to the human body, especially for the elderly. But falls tend to be infrequent, making it difficult to collect large amounts of data for research. In reality, there is a large gap between the amount of sensor data collected by falling activities and daily activities, which will lead to class imbalance. When using machine learning algorithms to detect falls, class imbalance will cause the performance of the classifier to be biased towards most classes and reduce the detection accuracy of a few classes. When faced with the problem of binary class imbalance, selecting an effective machine learning algorithm and resampling data can effectively improve the accuracy of classification. In this paper, an ensemble learning algorithm and clustering undersampling method are used for fall detection. The ensemble learning algorithm can reduce the impact of imbalanced datasets on the training model through multiple classifier iterations. Clustering undersampling method can change the dataset distribution and balance the number of positive and negative samples. The method in this paper is evaluated on the public dataset Sisfall. Compared with the traditional machine learning algorithms, the ensemble learning has higher accuracy and faster training speed. Combined with the clustering undersampling method, the method has a higher recall and precision.