Apple TV search gets personal: hybrid embeddings boost accuracy 8.6%

Apple has unveiled a novel personalization system for incremental video search on Apple TV, designed to deliver high-quality ranking with each keystroke—even when user intent is vague (e.g., 1-3 character prefixes). The approach, detailed in a paper accepted to RecSys 2026, learns two complementary item embedding spaces: TextEmb, a multilingual encoder fine-tuned via contrastive learning on co-engagement triplets, and IdEmb, an ID-based collaborative filtering model trained on interaction-derived positives. At serving time, user representations are built from recent watch history, and cosine similarities from both embeddings are fed into a pairwise XGBoost ranker to produce personalized results.

Offline evaluations using temporally held-out datasets showed a 2.99% improvement in NDCG@10 and 3.30% in MRR over a non-personalized baseline. The gains were most pronounced on ambiguous short queries (1-3 characters), where NDCG@10 lifted 8.63% versus just 1.46% for fully specified queries. Users with longer watch histories (51-100 items) saw a +4.37% lift, even though their baseline relevance was lower. A three-week online controlled experiment confirmed statistically significant gains: +1.14% tap-through rate, +1.23% conversion rate, and a 2.91% improvement in converted-item rank position. Ablation studies highlighted the trade-offs between semantic coverage and collaborative precision, with LLM-judged similarity labels used to reduce click/exposure bias.