Tutorials
All tutorials will be held on September 28th, with a duration of 3 hours each.
- Reward Optimizing Recommender Systems
About
Organized by: David Rohde
A quote often (mis)attributed to Einstein states: “If my life depended on solving a difficult problem in one hour, then I would spend fifty-five minutes defining the problem, and only five minutes finding the solution”. An implicit consensus developed in the early 2000s seemed to provide such a clear problem definition to the field of recommender systems. A recommender system should be evaluated by its ability to impact business metrics as observed at A/B test time. However, having this clear problem definition did not result in a well defined practice in service of optimizing this goal. Instead, recommender systems leverage an ingenious, but, ad-hoc combination of collaborative filtering, foundational models, click models, off-policy estimators and reinforcement learning. This tutorial will take a first principles approach to optimizing long term reward of a recommender system, showing how theoretical frameworks such as Bayesian inference, causal inference and reinforcement learning relate to pragmatic recommender systems techniques such as collaborative filtering, off-policy estimation, click models and generative AI. It will then finally reflect on the possibility that recent innovations in Generative AI may finally open the door to a future with truly reward optimizing recommender systems.
- Composing Agents, Compounding Risks: A Tutorial on Robustness and Alignment in Multi-Agent Recommender Systems
About
Organized by: Kurt Cutajar, Jas Kandola, Anjun Hu and Yashar Deldjoo
Multi-agent recommender systems are emerging as a new design paradigm in which LLM-powered agents plan over context, invoke tools, exchange intermediate outputs, and maintain memory across interactions. While these compositions create new opportunities for interactive and modular recommendation, they also introduce system-level risks that are easy to miss when agents are evaluated in isolation. As recommendation increasingly arises from composed multi-agent systems rather than single models, a shared vocabulary is still lacking for how risks propagate through agent interaction and how they should be evaluated. In line with RecSys 2026’s emphasis on recommender systems as systems, this tutorial addresses that gap through a risk taxonomy for multi-agent recommenders, an evaluation framework that scales from component to composition across offline and online tests, and mitigations across the system lifecycle.
- Recommender Systems and Sustainability: Bridging Algorithmic Design and Societal Impact
About
Organized by: Allegra De Filippo, Ludovico Boratto and Giuseppe Spillo
As outlined by the United Nations’ Sustainable Development Goals (SDGs), sustainability is a multidimensional concept aiming to meet present needs without compromising future generations. In this context, Recommender Systems (RS) play a dual role: they can serve as tools to nudge users toward sustainable behaviors (RS for Sustainability) and must themselves be designed to be energy-efficient and fair (Sustainability of RS). While these perspectives are crucial, current literature often addresses them in isolation. This tutorial provides a comprehensive and systematic review of this dual perspective, distinguishing between systems that promote SDGs and those that embed sustainability principles into their design. We offer a taxonomy classifying works across social, environmental, and economic pillars and outline a roadmap for developing holistic RS that use multi-objective optimization to balance performance with ethical and environmental costs.
- Recommender Systems in Delivery Platforms: Challenges, Solutions and Learnings
About
Organized by: Marcel Kurovski, Paavo Camps and Raghav Saboo
Delivery platforms introduce unique challenges for recommender systems, as they must optimize a digital experience under physical and operational constraints. Items are tied to specific locations with delivery radii, real-time inventory, and preparation times – signals absent in other recommendation domains. More fundamentally, these platforms are multi-sided marketplaces where the interests of consumers, merchants, couriers, and the platform itself are often part of large multi-objective optimization. Recommendations must navigate these competing objectives at scale and incorporate feedback that is both delayed and heterogeneous, extending beyond standard engagement signals to fulfillment outcomes such as wait times and item availability. This tutorial provides a practical overview of how recommendation systems are designed under these constraints. We cover the full surface landscape – from homepage discovery to item-level ranking – and address challenges including cold-start in a three-sided market, the glocal problem of globally trained models under hyper-local delivery constraints, and the alignment of offline metrics (MRR, NDCG, Recall) to online and business outcomes (conversion, gross order value, and retention). We present production recommender systems spanning the full recommendation stack, from representation learning and retrieval to ranking, page optimization, and generative methods, across both store- and item-level surfaces. We cover cross-domain store ranking across restaurants, grocery, and retail, as well as approaches for item ranking and product hierarchy structuring. Delivered by practitioners from Wolt and DoorDash, operating at scale in 40+ countries, the tutorial provides a system- and model-level understanding of recommender systems in delivery marketplaces.
- Continuous Automated Reproducibility: Practical Tools and Workflows for Repeatable and Reproducible Research
About
Organized by: Michael Ekstrand
While the RecSys community has been promoting and discussing the importance of reproducible research for many years, there remains a significant lack of practical resources to help RecSys researchers and practitioners, particularly those just getting started in the field, to actually build reproducible experiments and research workflows. This hands-on tutorial intends to fill that gap. Based on the idea of “continuous reproducibility”, I will present open-source tools and practical experimental design and implementation techniques to build recommender systems experiments that are reproducible from Day 1. This tutorial will be useful for researchers at various stages who want to make their work more reproducible, as well as to industry practitioners and researchers who want to build reliable, adaptable, and reproducible pipelines top train and evaluate recommender systems for production applications.
- Transformer-based Sequential Recommender Systems
About
Organized by: Jan Malte Lichtenberg and Aleksandr V. Petrov
Transformer-based models have become the cornerstone of sequential recommendation, yet they are often perceived either as rigid engineering recipes or as a collection of disconnected architectures. This tutorial demystifies these systems by centering on a provocative guiding question: “What is \emph{not} sequential recommendation?” By framing recommendation as an inherently temporal task, where user intent, catalogs, and context constantly evolve, we present Transformers as a unified and practical framework for modern Recommender Systems. The tutorial is structured into two blocks. The first block establishes foundations using SASRec as a running base model, covering item embeddings, causal self-attention, and the impact of positional information. The second block dives into advanced trade-offs, comparing causal and masked (BERT4Rec) modeling, scaling to large catalogs via quantization and Seamtatic IDs, and integrating content-aware representations (DenseRec). Designed for both researchers and industry practitioners, this session prioritises intuition and reproducibility. Participants will interact with a live playground to observe how model assumptions affect real-time recommendations. Attendees will leave with a clear understanding of when to use Transformer-based models, how to navigate their architectural trade-offs, and a suite of open-source materials for robust implementation.
- Beyond Searching a Village: Learning to Recommend Diverse and Successful Collaborative Teams
About
Organized by: Mahdis Saeedi and Hossein Fani
Team recommendation involves selecting skilled experts to form an almost surely successful collaborative team, or refining the team composition to maintain and/or excel at performance. To address the tedious and error-prone manual process, various computational approaches have been proposed, especially for the web-scale social networks and widespread online collaboration and diversity of interactions. In this tutorial, with a brief overview of pioneering subgraph optimization approaches and their shortfalls, we deliberately focus on the recent learning-based approaches, with a particular in-depth exploration of graph neural network-based methods. More importantly, and for the first of its kind, we then discuss team refinement, which involves structural adjustments or expert replacements to enhance team performance in dynamic environments. Finally, we discuss training strategies, benchmarking datasets, and open-source libraries, along with future research directions and real-world applications. Further resources are available at https://fani-lab.github.io/OpeNTF/tutorial/recsys26.
- Cognitive Biases in Recommender Systems
About
Organized by: Markus Schedl, Antonela Tommasel and Oleg Lesota
Extensive research in psychology and sociology has revealed a plethora of cognitive biases that describe systematic deviations from rationality and objectivity in humans’ cognitive processes and judgment or collective prejudices of a society that favor one group’s values, norms, and traditions over others. Common examples include confirmation bias, conformity bias, or cultural homophily. Such biases are also reflected in user-generated data (such as ratings or tags) and, as a consequence, manifest in user or item models that are leveraged by recommender systems. By picking up these biases, recommendation algorithms may create undesired effects such as stereotypical recommendations or even harmfully unbalanced results. In addition, cognitive biases influence how users interact with recommender systems. In this multidisciplinary tutorial, integrating research in psychology and computer science, we discuss a selection of cognitive biases that are particularly important in the context of recommender systems (RS). We provide an introduction to them based on empirical evidence from psychology, engaging the audience to experience the corresponding biases themselves. We discuss where they are evidenced in the ecosystem of RS and outline ideas for how their detriments can be alleviated, but also their potential benefits could be leveraged. Finally, we integrate a hands-on part where participants are able to (1) design psychology-informed user choice models based on the presented cognitive biases, (2) apply them in a provided feedback loop simulation setup with established recommendation algorithms, and (3) observe the effects of various cognitive biases on the evolution of produced user profiles and recommendations.
Tutorial Chairs
- Kim Falk, Binary Vikings, Dk
- Maurizio Ferrari Dacrema, Politecnico di Milano, Italy