Simulating Personalized Smart-Home Activity Datasets with Generative AI: A Case Study

To create, evaluate, and compare different technologies of home health monitoring, people need lots of sensor data that captures (1) residents' activities of daily living (ADL) like bathing, and (2) their usage of home appliances like watching TV. Unfortunately, such datasets of real-world monitoring are quite limited and scarce, due to issues like sensor cost, technique complexity, and deployment time. Existing simulators attempt to resolve these issues, by generating synthetic data based on predefined models or interactions with users. However, they give little consideration to personas, and rarely support personalized simulation based on humans' age, lifestyles, or health conditions. This paper introduces our novel investigation of using ChatGPT to create usable and shareable datasets of (1) human daily activities, and (2) their usage of home appliances. Specifically, there are two phases in our investigation. First, we described personas of home residents and layouts of home appliances, in order to use ChatGPT to generate data that mimic human behaviors and schedule their usage of appliances. Second, we analyzed and visualized the generated data to check whether the data is meaningful. Our results show great promise: the daily activities of different humans match the described personas, and the simulated appliance usage resembles the typical appliance usage in real-world settings. Our work may shed light on future directions of ADL simulation. It also facilitates studies on home health monitoring, disease diagnosis, and home automation.