Dipartimento di Ingegneria informatica, automatica e gestionale

Speaker: Daniele Malitesta (LUISS Guido Carli University)

Title: Diffusion-Based Multimodal Graph Machine Learning Under Unreliable Graph Data

Zoom link: https://uniroma1.zoom.us/j/81178556941?pwd=7qp2VjUNNTOtA0A5cy95eaQkN9bRG...

Abstract

Graphs are a natural representation for a wide range of real-world data, from social networks and e-commerce interactions to molecules and AI agent communities. Graph neural networks have become the standard model for learning from such non-Euclidean structures, and their extension to multimodal settings (where nodes, edges, or features carry information from images, text, or other modalities) has opened new opportunities across domains such as personalized recommendation, healthcare, and agent-based task planning. Nevertheless, multimodal graph data collected in real-world scenarios is rarely clean: it is often incomplete, noisy, or inconsistent, undermining the reliability of downstream graph learning models. This talk presents my recent and current research on multimodal graph machine learning under this broader challenge of unreliability, and discusses how graph generative models can offer a principled way to address it, with a particular focus on the personalized recommendation domain. Building on this, I will outline a future research agenda centered on graph diffusion models (a notable and recent example of powerful graph generative approaches) and their potential to tackle unreliable multimodal graph data. This agenda spans three directions: measuring unreliability in multimodal graph data; designing diffusion-based strategies to mitigate it, both as a standalone correction step and as an integrated component of downstream pipelines; and understanding the limitations of diffusion models themselves, including scalability and fairness concerns. I will conclude by discussing how these directions could generalize beyond recommendation to other domains, such as healthcare and the emerging area of graph-based task planning for AI agents (a direction I am currently exploring at LUISS).

Zoom meeting: https://uniroma1.zoom.us/j/81178556941?pwd=7qp2VjUNNTOtA0A5cy95eaQkN9bRG...

Bio

Daniele Malitesta is a postdoctoral researcher at LUISS Guido Carli University. Previously, he held a postdoctoral position at Université Paris-Saclay, and before that he completed his PhD in Computer Science Engineering at Politecnico di Bari. His research spans graph machine learning, personalized recommendation, and multimodal deep learning, with a recent focus on multimodal graph learning and generation under unreliable graph data. His work has been published in leading conferences, including AISTATS, The Web Conference, ACM Multimedia, SIGIR, and CIKM, where he has also been nominated outstanding reviewer three times, as well as in high-impact journals such as IEEE Transactions on Knowledge and Data Engineering and Expert Systems With Applications. Daniele is active in the research community, organizing venues such as the Learning on Graphs Conference and the DaQuaMRec workshop at ACM RecSys, and giving invited talks, tutorials, and summer school lectures. He recently joined the editorial board of the journal Intelligenza Artificiale and currently serves as guest editor for two Special Issues, at Data Mining and Knowledge Discovery and ACM Transactions on Recommender Systems.