A Comparative Study of Validation Methods for Sensor-Based Human Activity Recognition Using Deep Learning Models

With the rise of wearable sensors and smart devices, human activity recognition (HAR) has become a vital research area in ubiquitous computing. Although many studies report high accuracy using k-fold cross-validation, these results often do not reflect actual generalization interpretation due to subject-dependent data leakage, where models test on activities from the subjects they were trained. This study compares traditional k-fold cross-validation with leave-one-subject-out (LOSO) validation using the WISDM dataset, highlighting the importance of proper validation techniques in HAR systems. We implemented and evaluated five advanced deep learning models - convolutional neural network (CNN), long short-term memory (LSTM), bidirectional LSTM (BiLSTM), gated recurrent unit (GRU), and bidirectional GRU (BiGRU) - on windowed raw sensor data. Our experiments showed significant differences between validation methods. The BiGRU model achieved 97.91% accuracy with k-fold cross-validation and 98.02% with LOSO validation, while the CNN model achieved only 92.55% and 93.44%, respectively. These results underscore the impact of both model architecture and validation approach on performance. Our findings stress the need for subject-independent validation strategies like LOSO to develop truly generalizable HAR systems.