📝 Description

This project aims to develop inherently interpretable graph neural networks (GNNs) and spatio-temporal GNNs for tasks such as forecasting on networks (e.g., traffic, energy, climate, or mobility data). As a conceptual starting point, the thesis builds on ideas from B-cos networks, which enforce an alignment between inputs and weights so that the model can be summarized by a single linear transform that directly reflects task-relevant features, yielding transparent explanations without relying on post-hoc methods.

The core goal of the thesis is to transfer and generalize these interpretability-by-design principles from image-based convolutional networks to graph-structured and spatio-temporal data. The student will investigate how alignment-based layers or related mechanisms (e.g., constrained attention, concept-like intermediate representations, or linearized readouts) can be integrated into GNN architectures so that each node, edge, or time step comes with a clear, quantitative explanation of its contribution to the prediction. The focus is not limited to the original B-cos formulation: alternative inherently interpretable architectures and combinations with standard post-hoc explainability methods (e.g., saliency or attribution methods for GNNs) and our recently-developed Koopman interpretability tool for STGNNs can also be explored and compared.

From an application perspective, the student will work with spatio-temporal forecasting problems on graphs, such as predicting traffic speeds on road networks, power demand on grid nodes, or other phenomena evolving over irregular spatial supports. Using frameworks like Torch Spatiotemporal and standard GNN libraries (e.g., PyTorch Geometric), the project will involve (i) designing and implementing interpretable GNN/ST-GNN architectures, (ii) training them on real-world datasets, and (iii) evaluating both predictive performance and interpretability (e.g., faithfulness and human-aligned explanations). By the end of the project, the student will gain experience in modern GNNs, spatio-temporal modelling, and modern techniques for explainable AI on graphs, with potential impact in domains where transparency and trust in model decisions are crucial.

Data: Public benchmark datasets for spatio-temporal forecasting on graphs (e.g., traffic, energy, or mobility datasets), to be chosen based on the student’s interests and project focus.

📚 References

• B-cos Networks: Alignment is All We Need for Interpretability
• Interpreting Temporal Graph Neural Networks with Koopman Theory
• Torch Spatiotemporal
• Graph Deep Learning for Time Series Forecasting

📨 Contact

Michele Guerra (michele.guerra@uit.no), Filippo Maria Bianchi (filippo.m.bianchi@uit.no), Simone Scardapane (simone.scardapane@uniroma1.it)