How Alipay Boosted Ad CTR and CPM with Cold‑Start Fixes, Knowledge Transfer, and Real‑Time Learning

1. Background

As Alipay’s advertising business expands, new traffic scenes are added, introducing challenges such as cold‑start, low click‑through rate (CTR), and later CPM pressure. Algorithmic optimizations were applied: sample‑enhanced cold‑start, cross‑scene and user‑segmented knowledge transfer, and system latency improvements.

2. Cold‑Start Optimization

Initially the ad slot mixed with recommendation cards lacked ad samples, causing a cold‑start problem. By aligning trace_id between recommendation and ad logs, recommendation samples were used to augment ad training data, yielding an offline AUC gain of +0.01. As ad data grew, the benefit faded, prompting future work on joint modeling of recommendation and ad content.

3. Knowledge Transfer Learning

3.1 Cross‑Scene Transfer

Multiple‑scene/multi‑task models (SharedBottom, MMOE, PLE, etc.) were evaluated. While unified modeling offers data augmentation, distribution gaps cause gradient conflicts. Experiments with V0‑V3 configurations showed that fine‑tuning a pre‑trained multi‑scene model on the target scene (full‑parameter fine‑tune) improved AUC and business metrics (CPM, CTR).

3.2 User‑Segmentation Transfer

Using MAML, knowledge from non‑cold‑start units was transferred to cold‑start units. Tasks were defined by user‑group segmentation, enabling a modest AUC lift (+0.003) and notable business gains (CPM, CTR). The MAML inner‑outer loop updates are illustrated.

4. System Real‑Time Optimization

4.1 Feature Real‑Time

Real‑time features were selected based on offline importance (tree‑based gain, DNN weight magnitude). The top features were prioritized for streaming pipelines.

4.2 Model Real‑Time: Online Learning

4.2.1 Background

Traffic structure, user behavior, and supply shift hourly, making offline daily models insufficient. An online learning (ODL) framework streams samples to continuously update the CTR model.

4.2.2 ODL Model Optimization

Initial ODL suffered from over‑fitting and catastrophic forgetting. Strategies applied: fixing the embedding layer, sample replay (mixing historical offline data), model hot‑restart, and reduced learning rate. These restored stability and improved technical metrics (AUC +0.0076) and business metrics (CPM +3.55%, CTR +4.0%).

5. Conclusion

Over the past year, Alipay moved from unified multi‑scene modeling to targeted single‑scene solutions, achieving significant CTR, CPM, and revenue gains through cold‑start mitigation, knowledge transfer, and real‑time system upgrades. Future work includes joint modeling of heterogeneous recommendation and ad domains, deeper knowledge‑transfer techniques, and better handling of catastrophic forgetting in online learning.

References

Modeling Task Relationships in Multi‑task Learning with Multi‑gate Mixture‑of‑Experts, KDD'18.

Progressive Layered Extraction (PLE): A Novel Multi‑Task Learning (MTL) Model for Personalized Recommendations, RecSys'20.

One model to serve all: Star topology adaptive recommender for multi‑domain CTR prediction, CIKM’21.

AdaSparse: Learning Adaptively Sparse Structures for Multi‑Domain Click‑Through Rate Prediction, CIKM’22.

Model‑Agnostic Meta‑Learning for Fast Adaptation of Deep Networks, PMLR'17.

A Comprehensive Survey of Forgetting in Deep Learning Beyond Continual Learning, 2023.

Experience Replay for Continual Learning, NeurIPS'19.

Out‑of‑Distribution Generalization via Risk Extrapolation (REx), ICLR’21.