Multimodality in Recommender Systems: Does It Help, and Should We Expect an Answer?

Abstract

Multimodal recommender systems are gaining popularity for their promise of improved performance by integrating diverse data sources such as text and visuals. Yet, a central question remains: does multimodality truly help? In the first half of this talk, I will present findings from a structured evaluation that systematically benchmarks reproducible multimodal models against traditional baselines. In the second half, I will step back to situate these results in the broader landscape of recommender systems, emphasizing the task-specific nature of effectiveness and the need to consider user decision processes. Together, these perspectives address whether we should expect a universal conclusion about the benefits of multimodality in recommendation.

Bio

Dr. Aixin Sun is an Associate Professor with Nanyang Technological University (NTU) Singapore. His research interests include information retrieval, recommender systems, and natural language processing.  His work has earned several accolades, including the SIGIR 2025 Test of Time Honorable Mention Award, the Best Student Paper Award at the IEEE International Conference on Services Computing in 2020, and a Best Student Paper Honorable Mention at SIGIR 2016. Dr. Sun is an associate editor for ACM Transactions on Information Systems (TOIS), ACM Transactions on Recommender Systems (TORS), ACM Transactions on Intelligent Systems and Technology (TIST), and an editorial board member of the Journal of the Association for Information Science and Technology (JASIST). 

Sequential Recommenders and Multimodal Inputs: Mitigating Data Quality Issues in Industry-scale Recommenders

Abstract
Recommender systems research keeps advancing with increasingly sophisticated models, but their impact in practice is often constrained by the quality of the underlying data. In this talk, I will draw on experiences from industry-scale recommender systems to highlight two perspectives: first, why ID-based sequential recommenders are often more resilient to certain data quality challenges than traditional learning-to-rank methods; and second, how multimodal content inputs can mitigate item cold-start, a particular form of data quality issue.

Bio
Jan Malte Lichtenberg is a Principal Applied Scientist at Albatross AI, working on foundational sequential recommender systems. Previously, he worked at Amazon Music on learning-to-rank algorithms, counterfactual off-policy evaluation of ranking policies, and personalization of LLM-based recommendation systems. He holds a PhD in Computer Science from the University of Bath, UK.

Single-Branch Architectures for Recommendation

Abstract
Single-branch architectures have been proven effective in several multimodal learning applications. In this talk, I will describe the use of single-branch architectures for recommendation. I will first describe their use in multimodal recommendation, showing how they allow to address missing modality and cold-start scenarios. I will then describe the use of single-branch architectures in collaborative filtering, showing how they allow to reduce the number of model parameters without substantially affecting the quality of recommendations.

Bio
Marta Moscati is a PhD student at the Johannes Kepler University Linz and member of the Multimedia Mining and Search group. Her research interests revolve around the topics of multimodal recommender systems, multimodal learning, psychology-informed recommendation, and user modelling. Before joining the Multimedia Mining and Search group, she completed a PhD in Theoretical Particle Physics at the Karlsruhe Institute of Technology.