SPARC routes and compresses item attributes for context-aware generative recommendation

_reachsumit · x · 2026-07-29

SPARC is a sequence-aware progressive attribute routing and compression framework for generative recommendation.
The paper argues that industrial recommendation data contains heterogeneous attributes such as category, brand, price, behavior type, and timestamp, and that naïvely expanding or compressing them both has drawbacks.
SPARC first learns context-aware field representations for each attribute type.
It then routes original, contextual, and identity representations into multiple slots to preserve complementary information under a fixed capacity.
Finally, lightweight cross-item interaction compresses each historical item into a single token.
The authors position the method as an Alibaba-oriented approach to making generative recommenders more efficient and context-sensitive.

More from Research