Apple's Approach to Personalized Video Search

Apple ML Research · rss · 2026-07-16

Apple ML Research detailed a personalized ranking system designed for incremental video search on Apple TV. The use case is characterized by the need to re-rank results upon every keystroke, often with highly ambiguous search intents indicated by prefixes of just 1–3 characters.

Methodologically, they combine two types of signals: one is a text-based multilingual encoder called TextEmb, fine-tuned via contrastive learning on co-interaction triplets; the other is IdEmb, a collaborative embedding trained on interaction data. At serving time, these signals are aggregated to construct user representations, thereby enhancing personalized retrieval and ranking.

Original post →

More from Apps

Apps channel →