SparseCTR: Evolving sparse attention for long behavior histories¶
Core public-data sequence mechanism.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Institute of Software, Chinese Academy of Sciences. |
| First public release | 2026-01-25 (arXiv v1) |
| Author code | laiweijiang/SparseCTR. |
| Local adapter or method key | sparsectr |
| Local reproduction code | src/auto_research/reproductions/sparsectr/ |
Paper method and evidence¶
TimeChunking adapts to a user's inter-event intervals. EvoAttention combines global chunk, interest-transition and local-window attention; RelTemporal adds relative interval, hour and weekend biases.
flowchart LR
A[Chunk by individual timing] --> B[Combine global transition and local attention]
B[Combine global transition and local attention] --> C[Add relative time biases]Original paper figure¶
The original figure shows temporal chunking and attention. From the original paper; copyright remains with the authors.
Core equation¶
\[
H=\alpha_gA_{global}+\alpha_tA_{transition}+\alpha_lA_{local},\qquad A=\mathrm{softmax}(QK^\top/\sqrt d+B_{time})V.
\]
Reported results¶
Seven-day 1%-traffic experiment: CTR +1.72%, CPM +1.41%, history length 128→1024.
Local reproduction¶
MovieLens-100K full catalog: NDCG@10 0.05401→0.05519 (+2.19%). Three-seed metrics, mean, SD and 95% CI. No private logs, production models, live traffic, or serving; public-data measures do not establish online lift.