At Rise of AI 2023 in Berlin, Andreas Grün, head of technology for ZDF's digital media, and Leif-Nissen Lundbæk, co-founder and CEO of Xayn, describe building energy-efficient, privacy-preserving recommendation systems for German public broadcaster ZDF's streaming platform Mediathek. Grün explains that roughly 50% of Mediathek's viewing volume comes from algorithmically generated recommendations, built by a personalization team ZDF formed in 2018, and that ZDF developed its own 'PSM metrics' (covering coverage, novelty, diversity and serendipity) and a public transparency site to bridge public-service-media values with data-science metrics, plus a debiasing method that increased overall viewing while surfacing less popular content. Lundbæk describes Xayn's approach of small, specialized language models rather than large general ones, citing a joint test where a zero-shot model matched ZDF's custom-trained recommender's click rate while using 98% less energy, computable on a mobile phone instead of a data center. Both argue Europe's stricter data constraints can drive more efficient, more carefully engineered AI rather than only limiting it.
Key takeaways
- Andreas Grün says about 50% of ZDF Mediathek's viewing volume comes from algorithmically generated recommendations, built by a team formed in 2018.
- ZDF created its own 'PSM metrics', including serendipity, to bridge public-service-media values with data-science evaluation.
- A 2022 debiasing method reduced recommendations of popular 'hero content' without lowering, and even increasing, overall viewing.
- Leif-Nissen Lundbæk says Xayn builds small, specialized language models that can run on a phone instead of a data center.
- A joint zero-shot test matched ZDF's custom recommender's click rate while using 98% less energy.
Detailed notes
ZDF's Mediathek and its personalization team
Grün says roughly 50% of viewing volume on ZDF's Mediathek streaming platform comes from algorithmically generated start-page content, and about 20% of all available content is recommended and consumed daily, far more than editors could curate manually. ZDF built a dedicated personalization and automation team in 2018, staffed with editors, data engineers and data scientists, including people with backgrounds in fields like nuclear physics, to develop its own recommendation algorithms under the constraints of German public-service-media law.
Bridging legal values and data-science metrics
He describes a gap between the language of regulators and editors, who talk about values like impartiality and transparency, and the language of data scientists, who talk about information-retrieval metrics like NDCG. To bridge this, ZDF developed its own 'PSM metrics' built on established measures such as coverage, novelty and diversity, with serendipity, exposing users to different content and topics, treated as central to the public-service mission. ZDF also launched a public transparency website in November explaining which parts of Mediathek are personalized and which algorithms are used, though Grün notes it drew only about ten direct responses.
Debiasing recommendations without losing viewers
Grün cites a 2022 paper on debiasing recommendation algorithms, tested through an internal A/B test, that reduced how much popular 'hero content' was recommended in a test group without lowering overall viewing volume, and viewing volume actually increased because recommendations became more tailored to individual interest rather than general popularity. On sustainability, he lists techniques ZDF already uses: smaller models, less training data, online reinforcement learning instead of large-batch retraining, user segments instead of fully individualized recommendations, and caching and adaptive computation timed to lower GPU load.
Xayn's small-model approach to green AI
Lundbæk describes Xayn's approach of building small, specialized language models rather than relying on large general-purpose models like GPT, arguing that large models are built for processing comparatively little data while his company's models need to process millions of documents efficiently. He says the models are distilled to be small enough to run on a mobile phone instead of a data center, and compares Xayn's model, called Xenia, against OpenAI's Ada 2 embedding model for tasks like similarity search, citing a 98% reduction in energy consumption.
A joint experiment: 98% less energy, comparable results
In a joint test on ZDF's 'stage teaser' next-video recommendation, Xayn's zero-shot model, trained only on public internet data with no access to ZDF user data, matched the click-through rate of ZDF's own sequential model trained specifically on user interactions, while consuming 98% less energy. Both speakers argue this shows cost and energy efficiency can be improved directly at the algorithm level rather than by simply avoiding AI, in an example involving roughly 80 million documents processed.
Q&A highlights
Asked about user experience in ten years, Grün says the near-term focus is fully automating and personalizing Mediathek and expects conversational AI to play a growing role, though he calls ten years too far to predict. Lundbæk says he looks for the minimum data needed per use case rather than maximizing personalization, and both note ZDF already uses synthetic data to model distributions without exposing real user data, while treating cross-user comparison cautiously because, per internal studies, users dislike being recommended content just because 'similar' users watched it.
Quotes
We've got a lot of bias in our data, so we've got a lot of problems here, but that's really important to address them openly.
Andreas Grün
We actually have to think about really deeply how can we produce better models, how can we produce better systems.
Leif-Nissen Lundbæk
Questions answered in this talk
How much of ZDF's Mediathek viewing comes from AI recommendations?
According to Andreas Grün at Rise of AI 2023, roughly 50% of viewing volume on ZDF's streaming platform Mediathek comes from algorithmically personalized start-page content.
How can recommendation systems use less energy?
Leif-Nissen Lundbæk of Xayn describes using small, specialized language models instead of large general-purpose ones, reporting a joint test with ZDF where a zero-shot model matched a custom-trained recommender's click rate using 98% less energy.
Does reducing popular content in recommendations hurt viewing numbers?
Andreas Grün says a 2022 debiasing method tested on ZDF Mediathek reduced how often popular 'hero content' was recommended without lowering overall viewing volume, and viewing actually increased because recommendations became more tailored to individual interest.