Subtransfer Learning in Human Activity Recognition: Boosting the Outlier User Accuracy

Human activity recognition (HAR) has become one of the most active areas of sensory healthcare research due to its potential to improve the quality of life both for patients in healthcare settings who are susceptible to falls or undergo physical therapy and the public to promote a more active lifestyle. The state-of-the-art machine learning and artificial intelligence algorithms analyze sensory data from multiple users to achieve unprecedented performance in recognizing daily activities. Most research focuses on the average performance across the population of users however, it is evident that the performance distribution across different users is anything but uniform. In fact, the outlier users demonstrate performance degradations of up to 30% compared to median accuracy on any given dataset. This study investigates the impact of outlier users on activity recognition and proposes a novel approach called subtransfer learning to demonstrate that transfer learning principles can be applied within the same dataset when coupled with augmentation techniques. Our results show that subtransfer learning outperforms the source model (SM) on four datasets with different age groups and sensor placements, even when using a few additional samples. Specifically, in the most challenging dataset, the proposed approach outperforms the SM accuracy on low-accuracy subjects (80% compared to 60%) and median-accuracy subjects (98% compared to 91%) while maintaining the same levels for high-accuracy subjects. A statistical significance test further demonstrates performance improvements compared to both source and subject-specific models with p-values 0.009 and 0.0005, respectively, across a range of dataset with demographically diverse users and sensor locations.