Problem
CTR models must serve many predictions under memory and latency constraints. Compressing parameters can reduce storage yet still lose AUC or add computation, so parameter count alone is not enough.
Approach
The framework applies low-rank compression to embedding tables and MLP layers. Instead of only approximating weights, it locally compresses output features to preserve the behavior needed by the prediction model.
Evidence
On two academic datasets and one industrial benchmark, the paper reports 3–5 times smaller models with faster inference and higher AUC than the corresponding uncompressed models. These are results for the evaluated architectures and compression settings.
Scope
The work connects ranking quality to deployment cost. It does not imply that arbitrary low-rank factorization improves accuracy, or that benchmark inference gains translate directly into online business lift.
Sources
Bibliographic details
- Authors
- Hao Yu, Minghao Fu, Jiandong Ding, Yusheng Zhou, Jianxin Wu
- Venue
- KDD 2024
- Status
- Published