Generative AI For Rapid And Scalable Behaviour Modelling In Cities

Understanding human behaviours is critical to improving operation efficiency, individual and organisational productivity, public health management, and quality of life in cities. The proliferation of sensors and Internet of Things leads to new opportunities and challenges for human behaviour modelling and forecasting behavioural patterns at scale. Annotating behaviours of interest is expensive and often infeasible. This demands new ways for modelling behaviours at scale, moving away from fully-supervised learning approaches. Recent Generative AI approaches and Large Language Models has become a compelling choice for modelling mobility behaviours, especially in cases when training data is limited. I will present our self-supervised learning (SSL) approaches for multimodal sensor data, with different applications as examples, including human behaviour recognition and traffic flow forecasting. I will also talk about how we handle data heterogeneity and fusion of heterogenous time-series towards robust representation. I will also present a new versatile paradigm, leveraging Large Language Models (LLMs) for time-series forecasting, including for mobility and energy domain, using natural language prompts.