Research map / Architectures and efficiency / RNNs and normalization
APG: Adaptive Parameter Generation Network for Click-Through Rate Prediction
Bencheng Yan, Pengjie Wang, Kai Zhang, Feng Li, Hongbo Deng, Jian Xu, Bo Zheng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add APG to deep CTR models (WDL,PNN,FIBINET,DIFM,DeepFM,DCN,AutoInt) | 7 deep CTR models · MovieLens/Amazon/IAAC | AUC: avg 79.46 -> 79.70, 69.06 -> 69.34, 65.26 -> 66.15 | Improved |
| generate full per-instance weights (basic APG) (ablation) | WDL · 3 datasets avg | AUC: 71.18 -> 71.45 | Improved |
| low-rank parameterization of generated weights (ablation) | basic APG · 3 datasets avg | AUC: 71.45 -> 71.45 | No clear effect |
| share U,V and generate only center matrix S_i (ablation) | low-rank APG · 3 datasets avg | AUC: 71.45 -> 71.58 | Improved |
| add over-parameterization of shared U,V (ablation) | APG v4 · 3 datasets avg | AUC: 71.58 -> 71.70 | Improved |
| decomposed feed-forwarding and parameter sharing for efficiency (ablation) | basic APG · IAAC large MLP | training time / memory: time -38.7%, memory -96.6% vs Base | Improved |
| self-wise condition instead of group-wise or mix-wise (ablation) | APG · 3 datasets | AUC: e.g. IAAC 65.90 -> 65.94 over mix-wise | Improved |
| input aggregation instead of output aggregation for mix-wise (ablation) | APG mix-wise · 3 datasets | AUC: IAAC 65.70 -> 65.90 (attention) | Improved |
| APG instead of Star/MMoE coarse-grained allocation | Base WDL · MovieLens, Amazon, IAAC | AUC and memory: MovieLens fg 79.28 -> 79.58 (Star vs APG) | Improved |
| enlarge specific-parameter rank K from small to large (beyond 256 no extra gain) (ablation) | APG · 3 datasets | AUC | Mixed |
| use small P (2-16) instead of large P (32-512) (ablation) | APG · 3 datasets | AUC | Hurt |
| deploy APG in industrial sponsored search | online model · IndusData | AUC / CTR / RPM: +0.2% AUC, 3% CTR, 1% RPM | Improved |
Closest papers by meaning
- Generalize for Future: Slow and Fast Trajectory Learning for CTR Prediction (AAAI 2024 · similarity 0.87)
- From Feature Interaction to Feature Generation: A Generative Paradigm of CTR Prediction Models (ICML 2025 · similarity 0.82)
- DiffusionNAG: Predictor-guided Neural Architecture Generation with Diffusion Models (ICLR 2024 · similarity 0.80)
- AT4CTR: Auxiliary Match Tasks for Enhancing Click-Through Rate Prediction (AAAI 2024 · similarity 0.80)
- AdaLRS: Loss-Guided Adaptive Learning Rate Search for Efficient Foundation Model Pretraining (NeurIPS 2025 · similarity 0.80)
- Noise Consistency Training: A Native Approach for One-step Generator in Learning Additional Controls (NeurIPS 2025 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
