Personalizing Progressive Video Search: Hybrid Text and ID Embeddings

Progressive video search must produce high-quality rankings for every keystroke, but 1-3 character prefixes make user intent unclear. Apple TV's personalization system addresses this by combining semantic and collaborative signals.

Two embedding models are trained: (i) TextEmb - a multilingual text encoder fine-tuned with contrastive learning on co-engagement triplets, (ii) IdEmb - collaborative filtering-based embeddings learned from user-item interaction data. At inference time, a user representation is constructed from recent viewing history, and Text- and ID-based user-item cosine similarities are fed into an XGBoost ranker.

In offline evaluation, sessions with user history showed a 2.99% improvement in NDCG@10 and 3.30% in MRR. The effect is especially large for short queries: NDCG@10 improved by +8.63% on ambiguous 1-3 character queries, compared to only +1.46% for longer queries. The improvement grows with more user history—users with 1-5 history items saw +2.13%, while users with 51-100 history items saw +4.37%.

In a 3-week online A/B test, click-through rate improved by +1.14% and conversion rate by +1.23%. The average rank of converted items also improved by 2.91%. The work also includes a trade-off analysis between Text and ID embeddings and an embedding quality evaluation using LLM judgment.