Adaptive Parameter Generation Network for Click-Through Rate Prediction

Deep click‑through‑rate (CTR) models are widely deployed in recommendation, search and advertising, but they typically use a static set of parameters shared by all samples. This static regime cannot capture the heterogeneous feature distributions of different users or items, limiting model expressiveness and leading to sub‑optimal performance.

To address this, the paper proposes a high‑performance, high‑effectiveness module called Adaptive Parameter Generation Network (APG). APG dynamically generates model parameters for each sample via a parameter‑generation network conditioned on sample features. Three conditioning strategies are explored: group‑wise (generate parameters per sample group), mix‑wise (combine multiple factors such as user‑item pairs), and self‑wise (use the sample’s own hidden representation).

Because naïvely generating full‑size parameter matrices would be prohibitively expensive, APG incorporates several efficiency techniques: low‑rank parameterization to factorize weight matrices, decomposed feed‑forwarding to avoid costly matrix multiplications, parameter sharing to separate private (sample‑specific) and shared components, and over‑parameterization to increase capacity without extra inference cost. These designs reduce both computation and storage overhead while preserving or improving predictive accuracy.

Extensive experiments on public datasets (Amazon, MovieLens, IAAC, IndusData) and internal Alibaba search‑ad data demonstrate that APG consistently boosts AUC for a variety of baseline deep CTR models (WDL, PNN, DeepFM, DCN, AutoInt, etc.). Offline results show up to 0.2% AUC gain, while online deployment yields +3% CTR and +1% revenue per mille. Ablation studies confirm the contribution of each component (basic APG, low‑rank, parameter sharing, over‑parameterization). Complexity analysis shows APG can cut storage by up to 96.6% and inference time by up to 38.7% compared with standard deep CTR models.

The study concludes that dynamic, sample‑specific parameter generation substantially expands the expressive power of CTR prediction models and represents a promising direction for future adaptive deep learning systems.