Adaptive Silhouette Tree builds semantic IDs with per-node branching for generative recommendation

_reachsumit · x · 2026-11-11

A new arXiv paper proposes Adaptive Silhouette Tree (AST) for constructing semantic IDs in generative recommendation:

Key insight: existing residual/hierarchical vector quantization assumes tokens carry independent meaning per level, but semantic IDs are hierarchical paths whose token meanings depend on prefixes.
Method: AST builds semantic IDs as tree paths via top-down divisive clustering, dynamically choosing each node's branching factor by maximizing the silhouette coefficient, with constrained variants enforcing minimum probability mass to improve long-tail supervision.
Results: significant gains over strong baselines across five Amazon Review datasets of varying sizes and domains.

More from Research