SPEAR: Selection-aware personalized rewriting and community search¶
Core public-data mechanism, not industrial traffic.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Dewu. |
| First public release | 2026-08-03 (arXiv v1) |
| Author code | mallocagi1-cell/spear. |
| Local adapter or method key | spear |
| Local reproduction code | src/auto_research/reproductions/spear/ |
Paper method and evidence¶
Generic rewrites can win via path-score shortcuts. Dual embeddings and gradient isolation protect retrieval semantics; multiply rewrite confidence by item relevance, then use a dynamic selector for request-specific weight, scale and bias.
flowchart LR
A[Protect retrieval embedding] --> B[Gate rewrite confidence by relevance]
B[Gate rewrite confidence by relevance] --> C[Apply request-specific selection]Original paper figure¶
The original figure shows adaptive rewriting and retrieval. From the original paper; copyright remains with the authors.
Core equation¶
Reported results¶
On 100,000 industrial sessions the paper reports semantic similarity@10 +18.2 and click recall@10 +99.5. Online query-view CTR +0.259 and mean reading depth +0.733; fully deployed in Dewu community search since 2025.
Local reproduction¶
Execute dual recall/rank embeddings, confidence × fidelity gate, dynamic scale and original-query residual. MovieLens-100K 260 users/420 items, seed 42: shared transition+content versus SPEAR NDCG@10 0.0354→0.0449 (+26.95%). Metrics. Item content proxies products/posts; no Dewu rewrite generator, private sessions or online ranker.