The 18th edition of the ACM Conference on Recommender Systems (RecSys) took place from October 14th to October 18th 2024, in Bari, Italy. As the leading international conference on recommender systems, RecSys continues to serve as the premier platform for presenting cutting-edge research, systems, and techniques in the field. This year’s event brought together top researchers, practitioners, and industry professionals from around the world, creating a vibrant space for knowledge exchange and networking. The conference featured a variety of sessions, including keynote speeches, tutorials, workshops, and poster presentations, fostering collaboration and sparking discussions on the latest advancements and challenges in recommender systems.
Criteo is one of the most visible tech companies at the RecSys conference, and this year Criteo was no exception with Criteo making 4 oral presentations throughout the conference:
- Flavian Vasile presented Welfare-Optimized Recommender Systems at the SURE workshop
- Flavian Vasile presented Can LLMs reliably reason about user preferences for Recommendation? ar the KaRS workshop
- Otmane Sakhi presented Logarithmic Smoothing for Pessimistic Off-Policy Evaluation, Selection and Learning (which is also a NeurIPS spotlight, reach out to Otmane Sakhi or Imad Aouali if you are going to NeurIPS 2024)
- David Rohde presented Why the Shooting in the Dark Method Dominates Recommender Systems Practice as an Industry Talk
Takeaways from our Criteos
Highlights from Imad Aouali
RecSys 2024 provided a refreshing, close-knit experience, distinct from the larger conferences like NeurIPS, ICML, and ICLR. This smaller setting facilitated interactions with familiar faces and prominent companies, creating a conducive environment for valuable technical discussions. With a strong industry focus, the conference addressed challenges directly relevant to Criteo. The collaborative atmosphere encouraged deeper exchanges with peers who shared similar research interests, opening doors to potential partnerships and valuable insights.
I had the opportunity to assist with Otmane’s presentation of our work, Logarithmic Smoothing for Pessimistic Off-Policy Evaluation, Selection and Learning, which was an exciting experience. Beyond our presentation, I enjoyed diving into the diverse presentations and posters at RecSys 2024. The conference covered a broad array of topics in recommender systems, including off-policy learning, collaborative filtering, sequential recommendation, bias and fairness, and, of course, large language models (LLMs).
My favorite session was Thorsten Joachims’ keynote at the CONSEQUENCES Workshop, titled Towards Steerable AI Systems. Thorsten presented three main ideas: first, exploring how knowledge from interactive systems can transfer to generative AI (GenAI); second, acknowledging the known challenges in GenAI and revisiting early concepts that might be adaptable; and third, considering what we can do differently by integrating GenAI with interactive systems. These insights resonated with me, as I’m steering my PhD research towards leveraging my background in interactive systems to develop innovative GenAI algorithms, particularly in LLM and diffusion models fine-tuning.
Highlights from Martin Bompaire
Attending the RecSys conference was incredibly exciting for two main reasons. First, as a machine learning engineer responsible for the recommendation models running in production, I saw it as a valuable opportunity to connect with other industry practitioners. Many have faced similar challenges, but have likely taken different approaches, offering a chance to learn from their experiences. Second, the conference provided access to cutting-edge research, often more forward-thinking, and it could inspire us in shaping our long-term roadmap.
Several intriguing topics were discussed, but one that particularly stood out to me was the exploration of how to combine collaborative filtering and content-based signals. Typically, recommender systems build item similarities either from past user interactions (collaborative filtering) or from product content, such as titles or images. While collaborative filtering tends to offer higher-quality signals, it falls short when dealing with rare or new items. On the other hand, content-based signals are available for all items, even the newest ones. With the rise of large language models (LLMs), we can now extract more complex relationships from content, offering valuable signals even for items without much interaction data.
The following four papers propose different approaches to combining collaborative filtering with content-based signals.
- SiBraR (A Multimodal Single-Branch Embedding Network for Recommendation in Cold-Start and Missing Modality Scenarios): SiBraR maps both collaborative filtering and content-based embeddings into a shared space using contrastive loss. The model is designed to handle scenarios where either content or collaborative filtering signals are missing by masking certain inputs, ensuring reliable recommendations even with incomplete data.
- beeFormer (beeFormer: Bridging the Gap Between Semantic and Interaction Similarity in Recommender Systems): beeFormer leverages a text encoder to generate item embeddings based on their content, which is then fine-tuned using collaborative filtering data. The approach incorporates a loss function similar to EASE, effectively blending semantic similarity with user interaction data.
- FLIP (Fine-grained Alignment between ID-based Models and Pretrained Language Models for CTR Prediction): FLIP generates a hybrid dataset by creating text descriptions for each item that include both collaborative filtering attributes (e.g., user and item IDs) and content data (e.g., item titles). It uses pretrained language models to extract insights from this text and integrates it with attribute-based features.
- Semantic IDs (Better Generalization with Semantic IDs: A Case Study in Ranking for Recommendations): This paper introduces the concept of Semantic IDs, which represent each item using a small set of content-derived identifiers. These IDs are generated by quantizing the item’s content embeddings into discrete tokens whose combination form the Semantic ID. Unlike the random IDs typically used in collaborative filtering, Semantic IDs are directly related to the item’s content, enabling better generalization by capturing meaningful relationships between items based on their attributes.
We are also developing our own approaches to combine collaborative filtering and content data. Stay tuned, we might present it at RecSys 2025 in Prague!
Highlights from David Rohde
I always particularly enjoy RecSys not just for the amazing talks and papers, but because its an important forum for Criteo to share its industry perspective and to share our advances in building state of the art recommendation engines.
As usual the CONSEQUENCES workshop was a highlight and I was particularly pleased to see @Otmane Sakhi present Logarithmic Smoothing for Pessimistic Off-Policy Evaluation, Selection and Learning, but I also enjoyed several other talks particularly Ciarán Gilligan-lee’s keynote on causal inference (I am planning to discuss some points of agreement and disagreement with Ciarán at a future AI Conversation — stay posted!) as well as Thorsten Joachims wide ranging and inspiring keynote.
Flavian also demonstrated his great flexibility in speaking on two quite different topics including Welfare Optimizing Recommendation at the SURE workshop, and diving into the issue of Can large language models use reasoning to do better recommendation at the KaRS.
I was also very pleased to step onto the stage of the Teatro Petruzzelli to present my industry talk on Why the Shooting in the Dark Method Dominates Recommender Systems Practice. As I hoped the provocative framing stimulated a lot of discussions (and both agreement and disagreement!) at the banquet that night and between sessions at the remainder of the conference.
Highlights from Otmane Sakhi
After dedicating significant time to study large-scale recommender systems for my thesis, attending the RecSys conference in person was a thrilling first for me! This RecSys edition presented an excellent opportunity to reconnect with old friends, discuss industry challenges, initiate new collaborations, and gain valuable insights into the future of recommender systems research. I was truly impressed by how the main conference and various workshops showcased the diversity within the RecSys community, each approach offering unique solutions to the recommendation problem.
I had the opportunity to present our NeuRIPS spotlight paper, Logarithmic Smoothing for Pessimistic Off-Policy Evaluation, Selection and Learning, as a talk at the CONSEQUENCES workshop. Even if this work is of theoretical nature, it was great to discuss its potential applications with RecSys researchers.The CONSEQUENCES workshop consistently stands out for us, not only because Flavian is co-organising it, but mainly because it truly reflects our research team’s dedication to exploring the interactive nature of recommender systems. I was fortunate to wrap up my duties on the first day of the conference, allowing me to spend the rest of the event learning and engaging with researchers.
I was particularly drawn to the successful applications of off-policy approaches in large-scale production environnement, especially the pressing question of how to align offline and online results in interactive systems. The conference acknowledged the significant real-world impact of these methods, featuring a dedicated oral presentation track. Among the many impressive works, I want to highlight “Multi-Objective Recommendation via Multivariate Policy Learning” by the ShareChat team, first-authored by Olivier Jeunen, a former Criteo colleague. This paper applies off-policy methodologies to a predefined set of RecSys parameters, optimize them offline, ensuring positive outcomes in A/B testing post-deployment. The work answers practical questions and offers rich methodological insights that could inspire some of the bidding and recommendation projects we’re tackling within our teams.
And, with that, RecSys 2024 is done 😃
Our participation at RecSys 2024 provided invaluable opportunities to showcase our work, engage with fellow researchers and practitioners, and foster new connections within the recommender systems community. The positive interactions and enthusiastic reception of our presentations underscored the critical role of collaboration and knowledge sharing in advancing the field. These exchanges not only highlighted the impact of our research but also reinforced the importance of collective efforts in driving innovation and addressing emerging challenges.
