SIM: Search-based User Interest Modeling with Lifelong Sequential Behavior Data for CTR Prediction¶

Core mechanism on public data; online and local evidence are separate.

Paper information¶

Field	Details
Paper	CIKM 2020
Institution	Alibaba.
First public release	2020-06-10 (arXiv v1)
Author code	No official author code linked by the paper as checked on 2026-08-09.
Local adapter or method key	`sim`
Local reproduction code	`src/auto_research/reproductions/sim/`

Paper method and evidence¶

Use the candidate item as a query. The General Search Unit retrieves a relevant subsequence from lifelong history; the Exact Search Unit attends to that subsequence for the candidate.

flowchart LR
  Q[Candidate query] --> G[General Search Unit]
  H[Lifelong behavior] --> G
  G --> S[Relevant subsequence]
  S --> E[Exact Search Unit]

Original paper figure¶

The original figure shows general search and exact candidate attention. From the original paper; copyright remains with the authors.

Core equation¶

\[ S=\operatorname{TopK}_{i\in H}s_{GSU}(q,i),\quad h_q=\sum_{i\in S}\operatorname{softmax}(q^\top k_i)v_i. \]

Reported results¶

Main Alibaba display-ad traffic reported CTR +7.1%, RPM +4.4%, and histories up to 54,000 events.

Local reproduction¶

Shared transition + content is the baseline; SIM changes only the core mechanism. On MovieLens-100K full-catalog ranking, NDCG@10 0.0354 → 0.0471 (+32.97%). Fixed metrics.

auto-research reproduce --paper sim --dataset-dir data --seed 42

This exercises candidate-conditioned retrieval on public data, not private logs or online serving. The local relative change is not the original traffic lift.