Semantische IDs schlagen Video-IDs für Billion-Scale-Empfehlungssequenzen

The standard approach to short-form-video recommendation—modeling user watch histories as sequences of atomic Video IDs—hits two hard walls: those IDs carry no semantic signal, and the Transformer’s quadratic self-attention chews through memory long before you hit useful sequence lengths. A new paper (arXiv:2606.07546) deployed at billion-user scale replaces both bottlenecks with a pair of surgical fixes: Semantic IDs and a Global-Aware Compression Transformer (GACT).

Semantic IDs: Kill the Embedding Table Bloat

Traditional Video IDs are orthogonal tokens—each video gets its own embedding, the table grows with corpus cardinality, and the model sees no relationship between a cat video and a dog video. The authors replace these with depth-truncated, coarse-grained Semantic IDs derived from content features. Embedding size drops from corpus cardinality to something manageable. Cold-start content? No problem: shared semantic prefixes let the model generalize immediately. No retraining, no massive embedding tables.

Quadratic Self-Attention? Fold the Sequence Instead

Even with compact IDs, stuffing a million steps of user history into a standard Transformer is a non-starter. GACT introduces non-parametric temporal folding—chunks time steps into a compressed representation—and a unified global query that attends to the entire folded sequence. Offline profiling on production infrastructure shows an order-of-magnitude drop in peak memory and a drastic reduction in FLOPs. That memory headroom buys the ability to feed longer user histories without blowing the budget.

Online Impact: Satisfied Engagement, Not Just Click Metrics

The paper reports substantial online gains from large-scale A/B tests: satisfied user engagement and satisfied content consumption both improved. These are not trivial vanity metrics—they mean users watched more of what they actually wanted, not just what the thumbnail grabbed. The semantic-native approach also makes the system inherently more interpretable (shared prefixes signal genre or style) and easier to maintain (no billion-entry embedding tables to shard).

What this really means: the next wave of recommendation models will stop treating items as opaque IDs and start embedding content semantics directly into the sequence. The compression transformer architecture is just the lever that makes it practical. Expect every major short-video platform to follow this pattern within a year.

Source: Beyond Item IDs: Scaling Short-Form-Video Recommendation via Semantic-Native Long Sequence Modeling
Domain: arxiv.org