SAM: Post-Purchase Interest Suppression and Personalized Recovery¶

Core mechanism evidence. Paper and local evaluations are separate.

Paper information¶

Field	Value
Paper	arXiv 2607.12714
Affiliations	Alibaba Group
First public release	2026-07-14 (arXiv v1)
Authors' code	No official author code linked as of 2026-08-09.
Adapter / method	`sam`
Local code	`src/auto_research/reproductions/sam/`

Original paper¶

Background and contribution¶

Purchase satisfies a cluster of interests, so reinforcing pre-purchase clicks can repeat recommendations. Pointwise cross-attention finds history to suppress; long purchases predict replenishment. ASGU strongly suppresses just after purchase and restores attention near the cycle through a log mask.

flowchart LR
  H[Purchase and click history] --> I[Intent suppression]
  H --> R[Replenishment rhythm]
  I --> G[ASGU log-mask]
  R --> G

Key figure from the paper¶

Original Figure 1: SAM. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism¶

\[ m_{ij}=1-a_j\sigma\left(k\left(\alpha-\frac{\Delta t_{ij}}{c}\right)\right),\qquad \ell_{ij}=q_i^Tk_j/\sqrt d+\log\operatorname{clip}(m_{ij},\epsilon,1). \]

Paper results¶

Paper vs. DIN: CTR +1.1%, GMV +0.9%, bad cases −74.5%, and PPRR 15.5%→4.1%.

Local reproduction¶

Compare DIN-like scores to genre-based cycles and ASGU: Hit@10 −12.50%, NDCG@10 −6.60%. Intent/rhythm paths, ASGU, log-mask and TTNP run. Metrics.

auto-research reproduce --paper sam --seed 42

Reproduction limits¶

High-rating genre repetitions proxy purchase intervals; no replenishment goods or censored purchase logs. Negative local results do not extrapolate online.