Feb 23
Recombee Research 2024
Academic Foundations
Recombee has always been deeply connected to academia, with four of our six co-founders holding Ph.D. degrees. Over the years, our investment in research has grown alongside our company and the increasing demand for advanced features in the market.
This commitment is exemplified by the achievements of RecombeeLab, our joint research laboratory with the Faculty of Information Technology at the Czech Technical University in Prague. In 2024, RecombeeLab had another highly productive year.
Beyond this collaboration, we’ve built an internal team dedicated to transitioning innovations from academia into practical applications within our recommender systems. This focus has allowed us to effectively integrate cutting-edge research into products that deliver real impact for our clients, aligning with the latest AI trends.
Breakthroughs in Recommendation Systems
Introducing beeFormer
One of our standout innovations of 2024 is beeFormer, an open-source sentence transformer designed to meet diverse generative AI needs.
Presented at ACM RecSys 2024,[1] beeFormer combines transformer architecture with a unique methodology for distilling interaction data. This approach improves natural language model capabilities for recommendation tasks, such as item segmentation and semantic search, paving the way for greater precision and flexibility.
Understanding Trends and Regional Interactions
At WWW 2024 in Singapore,[2] we introduced a method to assess shifts in item popularity over time, helping decode users’ temporal behaviors and detect changing trends in catalogs.
Our regionalization-based interactions algorithm[3] identifies optimal warehouse locations and other logistics opportunities based on user-item interaction patterns.
Generalization and Cognitive Science in AI
Our research efforts extend to understanding how personalized machine learning models generalize. At ICML 2024 in Vienna, we presented a study analyzing the generalization mechanisms of deep factorization methods,[4] which are widely used in recommendation systems.
We’ve also pioneered cognitive science modeling within the recommender domain.[5] One of our studies explores users' decision-making processes when interacting with item lists. Using similar methodologies, we analyzed how large language models (LLMs), such as GPT models, perceive object concepts.[6] This research reveals how certain LLMs adapt to human-like reasoning after fine-tuning, aiding in model selection and ethical AI development.
The Impact of Our Research
Prestigious Publications and Presentations
In 2024, our work appeared in high-impact journals like IEEE Transactions on Neural Networks and Learning Systems and ACM Transactions on Intelligent Systems and Technology. We also presented papers at leading conferences, including ACM RecSys and ICML, one of the "big three" AI conferences alongside NeurIPS and ICLR.
Sponsorship of ACM RecSys 2024
As part of our dedication to the field, we proudly sponsored RecSys 2024. Our researchers were active participants, presenting papers, sharing innovations, and contributing to discussions that shape the future of recommendation systems.
What’s Next?
Exciting Opportunities in 2025
Research will remain central to Recombee’s mission as we continue to push the boundaries of AI. A key highlight of the year will be ACM RecSys 2025, hosted in Prague which is also home to Recombee’s headquarters. Our team is looking forward to contributing to the organization of the event and exploring new ideas to advance the future of recommendation technologies.
Collaborate with Us
At Recombee, we value partnerships with researchers and innovators. If you’re interested in our work or exploring potential collaborations, reach out! We’re always open to exchanging ideas and advancing research together.
Contact us at research@recombee.com.
Resources
[1] Vančura, Vojtěch, Pavel Kordík, and Milan Straka. "beeFormer: Bridging the Gap Between Semantic and Interaction Similarity in Recommender Systems." In Proceedings of the 18th ACM Conference on Recommender Systems, 2024
[2] Alves, Rodrigo, Antoine Ledent, Renato Assunção, Pedro Vaz-De-Melo, and Marius Kloft. "Unraveling the Dynamics of Stable and Curious Audiences in Web Systems." In Proceedings of the ACM on Web Conference, 2024
[3] Alves, Rodrigo. "Regionalization-based Collaborative Filtering: Harnessing Geographical Information in Recommenders." ACM Transactions on Spatial Algorithms and Systems, 2024
[4] Ledent, Antoine, and Rodrigo Alves. "Generalization analysis of deep nonlinear matrix completion." Proceedings of the International Conference on Machine Learning (ICML), 2024
[5] Alves, Rodrigo, and Antoine Ledent. "Context-Aware REpresentation: Jointly Learning Item Features and Selection From Triplets." IEEE Transactions on Neural Networks and Learning Systems, 2024
[6] Hrytsyna, Anastasiia, and Rodrigo Alves. "From Representation to Response: Assessing the Alignment of Large Language Models with Human Judgment Patterns." ACM Transactions on Intelligent Systems and Technology, 2024 (To Appear)
Next Articles
Key Trends in News & Media for 2025 and the Role of AI-Powered Solutions
The way news is produced, delivered, and consumed keeps shifting, and as we move through 2025, a few big changes have become impossible to ignore.
Feb 20
2024 Wrap-Up
As we wrap up 2024, let’s take a moment to celebrate the milestones we’ve achieved together. Your collaboration and feedback have been at the heart of our progress. Here’s a quick recap of last year’s highlights.
Jan 27
Celestial Tiger Entertainment launches new Chinese Movie app, CMGO, with DIAGNAL
With Recombee’s AI-powered recommendation engine working with DIAGNAL Enhance, CMGO serves up personalised experiences for each viewer, driving engagement for the service.
Nov 14, 2024
