Generative models are changing recommendation systems. They can add information about items and may even help create a whole recommendation page. At ML Prague 2026, Bogdan Bodnar and I gave a workshop about these ideas. I work at Sky on recommendations for streaming services such as Peacock and SkyShowtime. I am also a PhD candidate studying how to optimize a whole recommendation page. In this workshop, I shared what I have learned in both roles. We prepared slides and notebooks, which are available in the workshop materials.
We covered three topics. First, using generative models to add metadata to items. Second, using transformer models such as SASRec to learn from a person's interaction history. Third, generating a whole recommendation page. I planned to cover all three in our three-and-a-half-hour workshop, but I overestimated how much we could fit in. We did not get to the third topic.
Giving items richer descriptions
Recommendation systems use information about each item, such as its genre, tags, and description. This information can be missing or inconsistent. Generative models can help fill in the gaps by writing descriptions or adding tags that capture more about an item.
This is one way to use generative AI in a recommendation system without replacing the recommender. The new descriptions and tags become extra information for the model. We still need to check that this information is useful and accurate, and that it is worth the effort to create.
Transformers for sequential recommendation
The next topic was sequential recommendation. A person's recent activity can tell us what they are interested in now. Transformer models such as SASRec use self-attention to learn patterns in the order of a person's past interactions and use those patterns to recommend items.
We used examples to connect the model to recommendation data and experiments. Transformers are also useful outside chat and text generation. They can learn from user histories and improve rankings. For example, they can rerank search results while the existing search system continues to find the items.
From ranking items to generating a page
The third topic is the subject of my current PhD research. I am studying a transformer model that generates a whole recommendation page. Many streaming services show recommendations in rows, with each row built around a different theme. A good page needs to choose items and arrange them into rows that are relevant, varied, and easy to understand.
I wanted to end the workshop with this topic and show how generative models could help create a full page of recommendations. But the three hours went by quickly, and we did not reach it. The slides and notebooks are available in the workshop materials for anyone who wants to explore the topics further.
Two days at ML Prague
ML Prague had two full days of interesting talks. Many were about large language models and other generative models, which makes sense given how much attention these topics get today. I would have liked more talks about how the models work, how they are trained, and how we measure their quality, alongside talks about using them in practice.
Three talks stood out to me. Lucie Blechova from Similarweb showed how her team replaced expensive LLM inference in a large data pipeline with a fine-tuned embedding model. She also shared lessons about speed, cost, and quality. Tomas Dulka from Aisle spoke about evaluating AI agents that find and fix security problems in code. He explained why testing systems with unpredictable results takes more than a simple pass or fail. Manikanth Reddy Gonchi from Bloomreach described how they use a SASRec-style model to rerank the top 200 search results from an existing learning-to-rank system. I learned a lot from all three talks and from their examples of building and evaluating these systems.