PaperlayerView exact revision on arXiv

Source-linked AI summary

PinDCO: Whole-Page Aware Dynamic Creative Optimization at Scale

Yu Hao, Yuchun Li, Peimeng Sui, Meilin Liu, Tianyuan Cui, Hao Li, Zicong Zhou, Akanksha Baid

arXiv:2609.11943v1cs.IRcs.LG

TL;DR

Generative AI is rapidly expanding creative variants, creating a need for DCO that personalizes selection under latency, cost, and whole-page constraints. PinDCO combines component-aware creative scoring, exploration-driven data collection, pixel-aware adjustment, and efficient candidate serving. On Pinterest production traffic, it delivered a +3.09% ad CTR lift with positive whole-page metrics.

  • Problem

    Generative AI expands creative repositories and creates cold-start, latency, and serving-cost challenges for personalized creative selection.

  • Method

    PinDCO uses component-specific CCFN scoring conditioned on ad-level predictions, exploration–exploitation data collection, pixel-aware size adjustment, lightweight pruning, caching, and dynamic batching.

  • Results

    +3.09% ad CTR lift was observed on deployed Pinterest advertising traffic, together with positive whole-page metrics.

  • Takeaways & Limitations

    PinDCO provides a production DCO system for matching a growing repository of creative variants to audiences while accounting for Pinterest’s waterfall-grid whole-page effects.

  • Takeaways & Limitations

    Future work targets greater scoring expressiveness, richer creative signals, user-sequence modeling, and tighter integration between creative generation and selection.

Abstract

from arXiv · show

Recent advances in generative AI have substantially accelerated the creation of high-quality ad creatives, dramatically expanding the number of candidate variants per campaign. This shift increases the need for scalable dynamic creative optimization (DCO) systems that can match creatives to the most relevant audiences under stringent latency and cost constraints. We present PinDCO, a production DCO system for ad creative retrieval and selection on Pinterest, a billion-scale visual discovery platform. PinDCO is built around a Creative Component Fusion Network (CCFN) that performs dynamic creative scoring by modeling each creative component (e.g., image, title, layout) with a dedicated tower, using component-specific hyperparameters to account for differing modeling complexity. The component representations are fused to predict a creative-level score conditioned on the ad-level prediction, and we improve training data quality via an exploration-exploitation strategy. To account for Pinterest's waterfall grid layout, where a creative's rendered size affects nearby content and session-level engagement, we introduce a Pixel-aware Adjustment Module(PAM) that adjusts scores based on creative size to encourage efficient screen real-estate utilization and better whole-page outcomes. To support the large volume of creative candidates, we further employ a lightweight pre-selection model for early pruning, and optimize serving efficiency through caching and dynamic batching. Extensive offline analyses and online A/B experiments demonstrate the effectiveness of PinDCO, yielding a +3.09% lift in ad Click-Through Rate(CTR) with positive whole-page metrics. With the strong performance, we launched PinDCO in the Pinterest Ads platform.

1 Introduction

Pinterest’s expanding creative space requires DCO that personalizes variants while respecting whole-page layout effects, massive serving volume, and evolving assets. PinDCO combines component-aware scoring, pixel-aware adjustment, exploration, and efficient serving, achieving a reported +3.09% ad CTR lift with positive whole-page metrics.

  • Motivation: Pinterest’s waterfall grid makes creative height consequential because taller creatives consume pixels and push subsequent content downward.Columns have fixed widths but variable heights, creating cross-slot interactions that conventional slot-independent DCO does not capture.
  • Motivation: Generative image models continually expand creative repositories, creating cold-start and high-QPS serving challenges for personalized selection.Exploration traffic supplies less-biased feedback for new variants and helps adaptation to changing creative and user-preference distributions.
  • Approach: PinDCO uses a Creative Component Fusion Network with dedicated towers for components such as images, titles, and layouts, predicting creative effects relative to ad-level scores.Component-specific hyperparameters address differing modeling complexity and asynchronous convergence across towers.
  • Results: +3.09% ad CTR lift was observed after deploying PinDCO on Pinterest advertising traffic, alongside positive whole-page metrics.The system was launched on the Pinterest Ads platform after offline and online evaluation.
  • Approach: A Pixel-aware Adjustment Module modifies scores by creative size, while lightweight pruning, caching, and dynamic batching control serving cost and latency.The system parallelizes DCO with ad ranking and prunes candidates before expensive downstream inference.

2 Related Work

Prior work scores creatives from content, campaign, and empirical signals and increasingly personalizes selection online. PinDCO addresses scalability limits in these approaches by combining early lightweight selection with parallel fine-grained scoring.

  • 2.1 Offline Creative Quality Scoring: Creative scoring commonly uses component content features, component interactions, campaign setup, or empirical user responses to predict CTR and related objectives.Empirical responses are important because subtle creative differences can be difficult to infer from content alone.
  • 2.2 Dynamic Creative Optimization: Exploration–exploitation methods balance unbiased feedback collection against the cost of serving underperforming creative variants.Prior examples include multi-armed bandit approaches for carousel or creative optimization.
  • 2.2 Dynamic Creative Optimization: Personalized DCO matches variants to audience segments using real-time user and contextual signals rather than assuming one universally optimal creative.Selection is typically modeled online to adapt variants to audience differences and shopping-journey context.
  • 2.3 Serving Efficiency: Traditional recommenders use lightweight retrieval followed by heavier ranking, but injecting many creative candidates into ranking becomes difficult as variants per ad grow.CACS adds sequential creative selection, whereas Peri-CR parallelizes selection but still ranks every variant for every ad candidate.

3 Methodology

PinDCO places a creative optimization branch alongside the conventional retrieval-and-ranking pipeline. It expands ads into creative candidates, prunes them locally, applies CCFN and pixel-aware scoring, and improves efficiency through caching and parallel execution.

  • Architecture: PinDCO parallelizes creative optimization with the expensive ad-ranking stage so creative-model latency is hidden rather than added end to end.This design leverages ranking latency while preserving a higher-capacity creative scorer.
  • Creative Asset Manager: The creative asset manager expands retrieved ads into creative candidates using metadata from a key-value store, with caching reducing backend load.The retained metadata is intentionally minimal because of high QPS and large index volume.
  • Pre-Selection: A local lightweight pre-selection model removes low-opportunity creatives before CCFN, avoiding extra model-server RPCs and infrastructure overhead.Its features come directly from the creative index, and candidate retention can be tuned according to performance-cost trade-offs.
  • Pre-Selection: Pre-selection aggressiveness trades downstream performance opportunity against serving cost, with extreme cases allowing one candidate to bypass later stages.Retaining more candidates often improves downstream performance but increases cost.
  • Architecture: The pipeline runs CCFN for creative scoring, applies the Pixel-aware Adjustment Module for whole-page optimization, and uses a multi-armed bandit for exploration and exploitation.These stages are summarized in the PinDCO architecture and follow candidate expansion and pruning.

3.3 Creative Component Fusion Network

CCFN models creative components in separate towers, encodes their features, and fuses their representations into a creative delta added to the ad-level prediction. Component-specific features and hyperparameters support personalization and differing component complexity.

  • 3.3 Creative Component Fusion Network: CCFN learns component-specific representations for elements such as image, title, and layout before fusing them into a holistic creative representation.Its building blocks are feature encoding, component-specific towers, and component fusion.
  • 3.3 Creative Component Fusion Network: Feature encoding normalizes continuous inputs, embeds categorical or ID features, and groups inputs by creative component or a shared category.The resulting representations are consumed by the component towers.
  • 3.3 Creative Component Fusion Network: The fusion layer produces a creative delta score that is combined with the ad-ranking prediction, allowing CCFN to focus on creative-level differences.The summed logits are supervised with the final user-engagement label, while the base model mitigates ad-level confounding.
  • 3.3 Creative Component Fusion Network: CCFN incorporates content, user, and empirical features to represent creative appearance, personalization signals, and historical variant performance.Content features include image and text embeddings plus engineered visual attributes; user features include demographics and activity-derived signals.
  • 3.3 Creative Component Fusion Network: Distinct tower complexity and convergence rates motivate component-specific hyperparameters such as dropout, which stabilize training and improve creative selection.Images are described as higher-dimensional and more difficult to learn, creating risks of asynchronous underfitting and overfitting.

3.4 Pixel-aware Adjustment Module for Whole-Page Optimization

PAM adjusts creative scores to balance ad performance against the screen space consumed by creatives and their effects on other page content. It uses aspect ratio as a size proxy, applies a tunable penalty, and selects the penalty strength through offline replay.

  • 3.4 Pixel-aware Adjustment Module for Whole-Page Optimization: PAM penalizes larger creatives unless their incremental performance gains justify the screen space they consume.The objective accounts for cannibalization of other ads and organic content by optimizing effective pixel utilization across the page.
  • 3.4 Pixel-aware Adjustment Module for Whole-Page Optimization: The final score multiplies the CCFN output by an aspect-ratio penalty, with penalty strength controlled by tunable k.The penalty function is applied to the CCFN score and uses relative aspect ratio as its size input.
  • 3.4 Pixel-aware Adjustment Module for Whole-Page Optimization: The penalty function is monotonically decreasing, bounded between 0 and 1, and implemented with a tanh form for simplicity.Alternative penalty functions remain possible and are left for future work.
  • 3.4 Pixel-aware Adjustment Module for Whole-Page Optimization: Offline replay tunes k by simulating winner selection from logged prediction scores and aspect ratios, such as to reach a target average aspect ratio.Summary statistics from simulated winners guide the choice of penalty strength.

3.5 Exploration and Exploitation

PinDCO uses epsilon-greedy exploration and exploitation rather than always selecting the highest-scoring creative. This balances data collection against production constraints while feeding engagement signals into feature aggregation and training labels.

  • 3.5 Exploration and Exploitation: Epsilon-greedy exploration addresses selection bias from many new creatives, subtle performance differences, and shifting user preferences.The strategy balances performance needs with serving cost, latency, and observability overhead.
  • 3.5 Exploration and Exploitation: The policy identifies the highest-scoring creative using the Final Creative Score and incorporates exploration into selection.Creatives are represented as a candidate set C, with c* defined as the score-maximizing creative.
  • 3.5 Exploration and Exploitation: The selected winner is served to the user, and resulting engagement signals support empirical feature aggregation and training-data labeling.

3.6 Serving Efficiency

PinDCO improves serving efficiency for high-QPS creative scoring with local caching, request sharding, and dynamic batching. These optimizations reduce feature-store load and limit latency from slow candidates.

  • 3.6 Serving Efficiency: The serving system retrieves online features and runs CCFN inference for candidate creatives at high inference QPS.Requests contain hundreds of ad candidates, each with multiple creative variants.
  • 3.6 Serving Efficiency: Local caching reduces backend feature-store load by routing each creative ID consistently to the same model-server machine.The cache threshold is increased to retain more entries without degrading inference performance.
  • 3.6 Serving Efficiency: Dynamic batching processes candidate scoring requests in parallel instead of waiting for one large request dominated by its slowest candidate.Parallel processing improves machine utilization and minimizes latency.

4 Experiments

The experiments compare PinDCO with baseline methods, assess individual components, and report online A/B-test improvements. PinDCO was subsequently launched on the Pinterest Advertising Platform.

  • 4 Experiments: Online A/B experiments report significant metric improvements for PinDCO relative to baseline methods.Metrics are reported as relative changes with respect to the baseline.
  • 4 Experiments: The evaluation includes analyses of baseline comparisons and individual system components.
  • 4 Experiments: PinDCO was launched on the Pinterest Advertising Platform after the online experiments.

4.1 Offline Experiments

Offline experiments show that creative-aware models improve selection quality over the no-CR baseline, with CCFN achieving the strongest overall results. Ablations further show that exploration data and component-specific dropout each contribute to model quality.

  • Offline Results: 0.171% PR-AUC and 0.046% AUC-ROC gains make CCFN the best overall model against no-CR.Peri-CR improves by 0.163% PR-AUC and 0.043% AUC-ROC, while CCFN achieves the largest reported gains.
  • Ablation Study: Removing either exploration data or component-specific dropout reduces performance, confirming that both design choices improve CCFN quality.Table 2 evaluates each removal relative to no-CR while keeping training-data volume constant for the exploration comparison.
  • Ablation Study: Exploration traffic improves empirical signals by providing broader, less biased coverage for new and subtly different creative variants.The ablation analysis links this coverage to reduced selection bias and mitigation of cold-start effects.
  • Ablation Study: Component-specific dropout balances towers with different modeling difficulty and convergence rates, stabilizing optimization and improving fused creative scoring.Removing this mechanism worsens representation learning for component-specific towers.

4.2 Online Experiments

Online experiments show that PinDCO improves ad engagement while preserving whole-page outcomes, with serving optimizations and personalization mechanisms supporting scalable deployment.

  • +3.09% Ad CTR lift makes PinDCO the strongest online creative-selection method, outperforming Lightweight-Only (+1.49%) and Peri-CR (+1.70%) against no-CR.The comparison is reported in Table 3 and is consistent with offline findings.
  • +0.04% Successful Session lift distinguishes PinDCO from Lightweight-Only (-0.24%) and Peri-CR (-0.12%), which improve Ad CTR but reduce whole-page outcomes.Successful Session includes both ad and organic actions, capturing whole-page impact.
  • 87% higher P99 and 114% higher P90 latency without dynamic batching demonstrate its importance for preserving serving efficiency.Without batching, candidates are scored in one large request, increasing straggler risk and reducing processing parallelism.
  • Caching reduces latency by 9.6% at P99 and 12.8% at P90 by reducing backend feature fetches, without directly accelerating inference.The gains come from optimizing data access and cache-hit latency.
  • PAM retains a 9.5% CTR gain while reducing aspect-ratio growth to 5.22%, compared with 9.8% CTR and 7.21% without the penalty.The result supports limiting creative pixel footprint while preserving engagement gains.
  • Personalized layouts align with user context: multi-image impressions increase with shopping intent, while single-image layouts appear more often for lower intent and architecture or design interests.Figures 3 and 4 compare layout impressions across shopping-intent and query-interest buckets.
  • Normalized incremental CTR gain rises rapidly and then plateaus as more variants survive pre-selection, supporting lightweight pruning at an efficient operating point.The monotonic pattern indicates diminishing marginal returns after sufficient diversity is retained.

5 Conclusion

PinDCO combines scalable creative matching, component-aware scoring, pixel-aware whole-page adjustment, pruning, and exploration-based training for Pinterest Ads. It delivered a +3.09% platform Ad CTR lift and was launched, while future work targets richer modeling and tighter generation-selection integration.

  • PinDCO matches expanding creative repositories to appropriate audiences through pre-selection, CCFN scoring, PAM adjustment, and multi-armed-bandit exploration.CCFN models creative components with component-specific towers and hyperparameters, while PAM balances slot engagement with whole-page impact.
  • +3.09% platform Ad CTR lift supported launching PinDCO on the Pinterest Ads Platform.
  • Future work aims to improve scoring expressiveness through stronger architectures, richer creative signals, and user sequence modeling.
  • The paper also identifies tighter integration between creative generation and selection as a direction for using learned preferences to guide more diverse, engaging creatives.
Loading 2609.11943v1…