RCLRec: Reverse curriculum learning for sparse conversions¶
Core public-data mechanism, not production conversion training.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Alibaba International Digital Commerce Group. |
| First public release | 2026-03-30 (arXiv v1) |
| Author code | No public author code found as of 2026-09-05. |
| Local adapter or method key | rclrec |
| Local reproduction code | src/auto_research/reproductions/rclrec/ |
Paper method and evidence¶
Select short conversion-related subsequences as reverse-chronological decoder prefixes. Joint generation supplies intermediate supervision; quality-aware loss suppresses uninformative curriculum samples.
flowchart LR
A[Behavior history] --> B[Select conversion-relevant events]
B[Select conversion-relevant events] --> C[Reverse prefix]
C[Reverse prefix] --> D[Quality-aware joint generation]Original paper figure¶
Figure shows reverse curriculum generation. From the original paper; copyright remains with the authors.
Core equation¶
\[
\mathcal L=\mathcal L_{target}+\lambda q(C)\mathcal L_{curriculum},\qquad C=\operatorname{ReverseSelect}(H).
\]
Reported results¶
Online deployment reports ad revenue +2.09% and orders +1.86%.
Local reproduction¶
MovieLens-1M fixed full catalog: NDCG@10 0.05401→0.04768 (-11.71%). Three-seed metrics, mean, SD and 95% CI. Proxy targets do not transfer to production. No private features, production traffic or serving. These local public-data results do not reproduce online lift.