Incremental video search personalization with Hybrid text and ID embedding
Incremental video search requires a higher ranking after each keystroke, where the intent is often not well-defined (eg, 1–3 letter initials). We present a personalized search system for Apple TV that integrates relevant semantic signals during ranking. Our approach learns two embedding spaces: (i) a text-based multilingual embedder (TextEmb) fine-tuned on interactive triplets by cross-learning, and (ii) an ID-based interactive embedding model (IdEmb) trained on interactive goodies. At runtime, we construct user representations from recent viewing history and convolve the text- and user-item ID cosine-like objects into the binary XGBoost standard. We test the system with temporarily hosted offline datasets and a three-week controlled online trial. Offline, for sessions with user history, a personalized position improves NDCG@10 by 2.99% and MRR by 3.30% over a non-personalized base. Unfortunately, the fragment analysis shows that personalization is more necessary in incremental searches, where the intent is still being formed: for vague prefix queries (1–3 letters), the NDCG@10 lift is +8.63%, compared to only +1.46% for long, fully specified queries. Users with a long viewing history benefit more from personalization than new users: NDCG lift increases from +2.13% for users with 1–5 history items to +4.37% for users with 51–100. This large lift occurs even though baseline compatibility is low in long history clusters (NDCG@10 drops from 0.733 to 0.680), indicating that personalization adds significant value when the default rate is performing below expectations. Online, the treatment yields statistically significant gains of +1.14% tap rate and +1.23% conversion rate, with a 2.91% improvement in converted item rate. We also analyze the trade-off between semantic and shared embeddings by using separate extraction of each signal, and evaluate the quality of embeddings in a captured corpus with similar labels judged by LLM to reduce click/reveal bias.
