Abstract:
The distinction between subject-dependent and subject-independent performance is ubiquitous in the Human Activity Recognition (HAR) literature. We test the hypotheses that HAR models achieve better subject-dependent performance than subject-independent performance, that a model trained with many users will achieve better subject-independent performance than one trained with a single user, and that one trained with a single user performs better for that user than one trained with this and other users by comparing four algorithms' subject-dependent and -independent performance across eight data sets using three different approaches, which we term person-independent models (PIMs), person-specific models (PSMs), and ensembles of PSMs (EPSMs). Our analysis shows that PSMs outperform PIMs by 3.5% for known users, PIMs outperform PSMs by 13.9% and ensembles of PSMs by a not significant 2.1% for unknown users, and that the performance for known users is 20.5% to 48% better than for unknown users.