A Residual Network with Focal Loss to Handle Class-imbalance Problem on Nurse Care Activity Recognition

Human Activity Recognition (HAR) is a remarkable basement for the development of the generalized and robust machine learning modeling, as this domain typically has to deal with high-dimensional sensor data that are characterized by a large mutability (e.g., due to the diversity of the user's actions or as the aggregate of noise). However, unlike other domains, HAR is a very challenging task because of the large scale of class-imbalance, lack of a good benchmark dataset for HAR, and different experimental setups (e.g., several types of sensors or changes in their locations). Highly imbalanced data forces the model to show large bias towards the majority class and thus, minority classes are almost ignored. On the “The 2nd Nurse Care Activity Recognition Challenge Using Lab and Field Data” dataset, we present a deep learning paradigm that is compiled with multi-class focus loss and residual block network. In our proposed model, the residual block network prevents the model from vanishing gradient problems in a deep learning network and focal loss overcomes the crucial class-imbalance problem. On the test data, we achieved a promising accuracy of 34.47%, beating the previous state-of-the-art algorithms.