Effective Knowledge Transfer for Multi-Task Recommendation Models¶
Core transfer mechanism on public data.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2605.05730 |
| Institution | Huawei Technologies. |
| First public release | 2026-05-07 (arXiv v1) |
| Author code | No official author code linked by the paper as checked on 2026-08-24. |
| Local adapter or method key | ektm |
| Local reproduction code | src/auto_research/reproductions/ektm/ |
Paper method and evidence¶
Transfer CTR representations into multiple conversion towers according to CVR-task similarity. A hard-example-aware loss limits negative transfer while keeping multi-task serving cost manageable.
flowchart LR
C[CTR representation] --> T[Similarity-weighted transfer]
T --> V[Multiple CVR towers]
V --> H[Hard-example-aware loss]Original paper figure¶
The original figure shows task-aware transfer. From the original paper; copyright remains with the authors.
Core equation¶
Reported results¶
Section 4.4.2/Figure 3 reports eCPM +3.93%, with traffic expanding 5%→20%→50%→100%. This production metric is not the local outcome.
Local reproduction¶
NDCG@10 0.0540→0.0599 (+10.98%), Hit@10 0.1091→0.1182. Same MovieLens-100K 220 users/360 items, full catalog, split and seed 42, using a shared transition + content baseline. Validation selects only blend weight; test does not tune. The local model has independent state and scoring. No private features, production checkpoint or online serving; retain negative results.