• Course Instructor: Daniele Zambon -  IDSIA (Istituto Dalle Molle di studi sull’intelligenza artificiale), USI (Università della Svizzera italiana). 

      Course Title: Graph Deep Learning for Time Series and Spatiotemporal Data

      When:  22 - 25 March and 13 - 15 April


      Aims of the course

      The course aims to provide a coherent and structured introduction to recent advancements in time series and spatiotemporal data processing enabled by graph deep learning. Students will acquire a rigorous methodological framework for modeling time series and spatiotemporal data using relational representations and graph neural networks. A second objective is to equip students with practical tools and guidelines for applying this framework across a broad range of applications, while addressing the challenges posed by real-world environments.

      Content(sketch):

      In many application domains, including smart cities, environmental monitoring, energy systems, and finance, data are generated across time and space by interrelated entities. The resulting spatiotemporal data form collections of time series that correlate due to spatial proximity, physical interactions, shared drivers, or more general functional dependencies. The course begins by formalizing this setting and positioning it within the broader landscape of multivariate and spatiotemporal modeling. While deep learning has proven effective for modeling complex temporal patterns, Graph Deep Learning (GDL) is presented as a framework that extends these capabilities by providing scalable tools for learning from spatially related data. Within this framework, the course introduces spatiotemporal graph neural networks (STGNNs), their core components, and a general recipe for designing models tailored to forecasting and related tasks. The second part of the course addresses practical challenges, including learning relational structure when it is not explicitly available, scaling to large systems, handling irregular or missing data, and assessing the quality of model predictions. Throughout the course, hands-on sessions expose students to open-source software libraries and demonstrate how to build and train spatiotemporal models. Key limitations of current methods and emerging research directions are also discussed to foster analytical thinking and prepare students for advanced study or applied work in the field.

    • Course Timetable

      • 23/03/2026, 16.30 - 18.30, Room: 1BC50
      • 24/03/2026, 16.30 - 18.30, Room: 1BC50
      • 25/03/2026, 9.30 - 11.30, Room: ROOM B PIOVEGO
      • 13/04/2026, 16.30 - 18.30, Room: 2AB40
      • 14/04/2026, 16.30 - 18.30, Room: 2AB40
      • 15/04/2026, 9.30 - 11.30, Room: 2AB40
    • Course materials