A fundamental step towards artificial intelligence is to provide
learning processes with the ability to constantly adapt to
continuously evolving environments and to deal with the
ever-increasing quantity of data that these environments generate.
This is the case in sensor-rich environments that characterize the
Internet of Things applications such as human activity recognition
(HAR). In this contribution, we propose to improve HAR systems with
domain knowledge composition. Three models encoding different domain
knowledge are used: (1) a model of the influence of data sources on
human activities, (2) a model of the interactions between data
sources, and (3) a model of the transitions between human activities.