SPARC: Sequence-aware Progressive Attribute Routing and Compression Framework for Generative Recommendation
Chang Liu, Changfa Wu, Hui Qian, Binbin Cao, Jian Wu, Yuliang Yan, Han Zhu, Bo Zheng
cs.IR
2026-07-28
SPARC contextualizes then route-compresses multi-field interactions for generative recommendation without lengthening the backbone; +6% hit rate on Taobao, +70% on Amazon.
Generative recommendation encodes each item as a discrete Semantic ID (SID) and autoregressively predicts the next item from a user's SID history, the way a language model predicts the next token. The trouble is that a real interaction carries far more than an item ID: category, brand, seller, price, inventory, behavior type, timestamp.
Getting those fields into the generative backbone has two bad options. Fully expand them and each field becomes its own token, so L interactions balloon from L tokens to L times F, which self-attention's quadratic cost cannot absorb. Or compress them into a single representation up front, which mashes the information together before anything knows what matters, dropping context-dependent signal (the same brand is predictive when a user keeps buying that brand, irrelevant for another user).
SPARC, from Alibaba, follows the principle "contextualize before compressing," in three stages.
Stage one, FCM, models sequential dependencies per field type. All "brand" fields along a user's history form one sequence, passed through a lightweight encoder to get a context-aware brand representation. Each field now has both its raw value and a contextual version.
Stage two, CAR, is the core. Fields split into two groups: SID fields (item identity) are preserved as-is and never mixed; side fields (category, price, behavior type) go through dynamic routing. A set of learnable slots computes how much each side field routes into each slot, and the same item gets different routing under different user histories, which is the context-dependent part. Each interaction ends up as a fixed number of tokens (SID tokens plus two side tokens), so the backbone input length does not grow.
Stage three, STC, fuses those intermediate tokens at the sequence level, using a residual gate initialized small to introduce cross-token interaction gradually rather than disrupting the compressed representation space.
On industrial Taobao data (21M users, 0.27B items, 2.6B interactions), SPARC layers on the RankGR backbone. Click hit rate HRclick@20 rises from RankGR's 0.1568 to 0.1669 (+6.4%); HRclick@1000 from 0.5777 to 0.5883. On the public Amazon Beauty and Toys sets the gains are larger (Beauty HR@20 from 0.0466 to 0.0794, about +70%), because sparser data benefits more from adaptive information retention.
The ablation pits SPARC against several static compressors (QFormer, MLP, Modulated); SPARC beats them all. The win comes from deciding what to retain based on context, not from stacking more compressor capacity.
| Metric (Taobao) | RankGR | SPARC | Gain |
| HRclick@20 | 0.1568 | 0.1669 | +6.4% |
| HRclick@1000 | 0.5777 | 0.5883 | +1.8% |
Backbone input length is a hard constraint in generative recommendation, and this work fits context-dependent multi-field signal into that fixed budget without lengthening the input. For practitioners in industrial recommendation it is a deployable change: it sits on top of an existing generative backbone, and the engineering is concentrated in the representation-compression stage. The learned routing patterns also match intuition (one slot tracks behavior type and recency, another tracks seller and brand).
The token budget per interaction is fixed (2 SID plus 2 side tokens), and the authors concede adaptive budget allocation is future work. Only nine fields are used; richer behavioral features are untested. Gains on Taobao-scale data are relatively modest, which the authors attribute to sheer interaction volume already easing part of the representation bottleneck, so extrapolating to less data-heavy settings deserves caution.