Using semi-Markov models for generating, validating, and analyzing artificial smart home processes

Improvements in medicine have led to huge increases in longevity, with a corresponding rise in demands for healthcare. Such changes have been accompanied by major developments in sensor technologies and smart homes, which can generate event data arising from a person’s Activities of Daily Living (ADLs). By identifying these underlying activities, technology can support older people in their homes through reminding them of routine activities, detection of anomalous behavior, and sending alarms to care providers. However, the lack of suitable large datasets of ADL process executions, for testing algorithms at scale, has been a barrier to the development and deployment of such solutions. ADL event data shares characteristics with datasets found in organizational task mining, characteristics such as relatively low volume and rich human-environment interactions. We here draw upon techniques from organizational process and task mining involving a semi-Markov modelling approach to characterize ADL sequences and activity durations and use these models to generate large artificial smart home process datasets. The methodology is evaluated on a public domain sensorized and annotated smart home dataset. Outputs are validated qualitatively through an analysis of the state transition matrices and graphs of activity duration, and quantitatively using both an entropy-based approach to compare duration distributions and conformance analysis techniques from process mining, demonstrating the viability of our approach at scale.