Personalizing Incremental Video Search with Hybrid Text and ID Embeddings

Details

Apple's ML Research team published a RecSys 2026 paper describing how they personalized Apple TV's incremental search — the query-as-you-type experience that must rank results after every keystroke, often from just 1–3 characters of input where user intent is highly ambiguous. The system learns two complementary embedding spaces: TextEmb, a multilingual encoder fine-tuned on user co-engagement triplets via contrastive learning, and IdEmb, a collaborative embedding model trained on interaction-derived positives. At serving time, recent watch history constructs per-user representations, and text- and ID-based user–item cosine similarities are injected as features into a pairwise XGBoost ranker. Offline evaluations on temporally held-out data show NDCG@10 gains of +2.99% and MRR gains of +3.30% overall. The effect is largest exactly where it matters most: on ambiguous 1–3 character prefix queries, NDCG@10 improves +8.63% versus only +1.46% for longer, fully specified queries. Users with richer watch histories (51–100 items) also benefit more (+4.37% NDCG lift) than new users with 1–5 items (+2.13%), and the personalization gain is highest in cohorts where the baseline ranker performs worst. A three-week online A/B test confirmed real-world lift: +1.14% tap-through rate, +1.23% conversion rate, and a 2.91% improvement in converted-item rank position — all statistically significant. The paper also runs ablations isolating each embedding type and evaluates embedding quality on LLM-judged similarity labels to reduce click/exposure bias.