Source-linked AI summary
Optimized Cost per Click in Taobao Display Advertising
Han Zhu, Junqi Jin, Chang Tan, Fei Pan, Yifan Zeng, Han Li, Kun Gai
TL;DR
Traditional fixed bids offer only coarse matching between traffic quality and advertiser value, while Taobao must also protect user experience and platform revenue. The paper proposes OCPC, which adjusts bids at page-view granularity and optimizes these interests jointly. Production online experiments report stable improvements across advertiser, platform, and overall metrics.
Problem
Fixed bids cannot finely match continuously varying traffic quality, while Taobao must balance advertiser demands with user experience and platform revenue.
Method
OCPC adjusts each page-view bid toward traffic quality while maximizing a composite score for user experience, advertiser interests, and platform revenue under unchanged eCPM sorting.
Results
67% of campaigns with more than 5 weekly conversions improved both GPM and ROI, while online experiments reported stable improvement across advertiser, platform, and overall metrics.
Takeaways & Limitations
OCPC was automatically applied to all mobile production traffic for Item CPC Ads and could also be selected by advertisers for Banner CPC Ads.
Takeaways & Limitations
The ROI optimization uses a fixed lower-bound threshold to avoid some advertisers receiving little traffic, and the cost model supposes a click cost equal to the advertiser bid despite actual GSP pricing.
Abstract
from arXiv · showhide
Taobao, as the largest online retail platform in the world, provides billions of online display advertising impressions for millions of advertisers every day. For commercial purposes, the advertisers bid for specific spots and target crowds to compete for business traffic. The platform chooses the most suitable ads to display in tens of milliseconds. Common pricing methods include cost per mille (CPM) and cost per click (CPC). Traditional advertising systems target certain traits of users and ad placements with fixed bids, essentially regarded as coarse-grained matching of bid and traffic quality. However, the fixed bids set by the advertisers competing for different quality requests cannot fully optimize the advertisers' key requirements. Moreover, the platform has to be responsible for the business revenue and user experience. Thus, we proposed a bid optimizing strategy called optimized cost per click (OCPC) which automatically adjusts the bid to achieve finer matching of bid and traffic quality of page view (PV) request granularity. Our approach optimizes advertisers' demands, platform business revenue and user experience and as a whole improves traffic allocation efficiency. We have validated our approach in Taobao display advertising system in production. The online A/B test shows our algorithm yields substantially better results than previous fixed bid manner.
1 INTRODUCTION
Taobao’s CPC display advertising system must match bids to continuously varying traffic while balancing advertiser outcomes, user experience, and platform revenue. OCPC addresses this with page-view-level bid optimization, and production experiments evaluate its effectiveness.
- Advertising setting: Taobao display advertising serves banner and item CPC ads on its mobile app, with item ads appearing among recommendation spots.The system focuses on CPC display advertising for Taobao mobile app traffic.
- Problem: Taobao must balance advertisers’ interests with platform ecological requirements, including user experience, GMV, and advertising revenue.GMV can benefit advertisers because many advertisers are also Taobao sellers, while media requirements constrain advertising spots.
- Pricing choice: CPC is adopted because it lets small and medium-sized advertisers control click costs while avoiding the higher risk of CPM and the weaker traffic liquidation efficiency of CPS.Taobao’s integrated data ecology and standardized e-commerce advertising process support CPC effectiveness.
- Problem: Fixed bids for user groups and ad slots provide only coarse-grained traffic differentiation, despite advertisers’ need to match bids with finer-grained traffic quality.A single fixed bid must handle continuous internet traffic with different quality levels.
- Proposed approach: OCPC adjusts bids toward page-view traffic quality while optimizing a composite score for user experience, advertiser interests, and platform revenue.The approach keeps the eCPM sorting mechanism unchanged and is designed to adapt flexibly to changing business needs.
- Evaluation: The paper reports comprehensive offline and online experiments to verify OCPC and its overall optimization of advertiser interests, user experience, and platform revenue.The evaluation includes model effectiveness estimation, offline experiments, and online A/B testing.
2 SYSTEM ARCHITECTURE
Taobao’s display advertising architecture processes each page-view request through coordinated components from request handling and targeting to candidate-ad search. The system architecture also incorporates the proposed bid optimization strategy.
- Architecture overview: The system architecture describes information flow from an initial page-view request to the final ad impression.The sequence of system components is presented through the star-schema architecture and proposed bid optimization strategy.
- Request coordination: The Front Server receives a page-view request and passes it to the Merger Server, which coordinates communication with other system components.The Merger Server acts as the central coordinator during the ad-serving process.
- Targeting and matching: The Merger Server requests user-targeting analysis from the Matching Server, which returns feature tags based on advertiser targeting requirements.These tags are then delivered through the Merger Server for candidate-ad retrieval.
- Candidate retrieval: The Search Node Server uses the delivered feature tags to search for candidate ads together with their bids.This connects targeting requirements with the candidate-ad search stage.
3 OPTIMIZED COST PER CLICK
OCPC adjusts campaign bids within advertiser-specific feasible boundaries and selects optimized bids to balance advertiser ROI and traffic quality with platform objectives such as GMV and advertising revenue. The ranking procedure preserves eCPM sorting while using composite objectives to choose ads and bids.
- Optimization scope: The framework adjusts each eligible campaign’s preset bid for a page-view request to satisfy advertiser demands and platform objectives.The typical setting optimizes ROI and quality traffic for advertisers while optimizing GMV and platform revenue.
- Optimization scope: Expected GMV per click is p(c|u,a) × v_a, while advertiser ROI depends on expected conversion rate, predicted pay-per-buy, and bid.The framework treats conversion probability as conditioned on the user and clicked ad, and assumes the bid is the click cost for deriving expected ROI.
- Optimization scope: OCPC raises bids for traffic whose conversion probability exceeds the crowd average and depresses bids for lower-quality traffic, subject to an ROI constraint.The optimization boundary uses lower and upper bid limits; a fixed threshold can prevent advertisers from receiving too little traffic.
- Ranking: The ranking objective can combine GMV and advertising revenue through f(·), with α controlling the trade-off between these platform goals.The selected ad maximizes the composite objective within the feasible bid region, while other candidates’ bids are adjusted so the selected ad has the largest eCPM.
- Ranking: For each candidate, feasible bid bounds induce lower and upper optimized rank-score bounds, allowing the system to select bids while retaining eCPM ranking.The rank score is pCTR multiplied by the optimized bid, and the selected ad must remain the highest-ranked under eCPM sorting.
- Algorithm details: Calibration reduces the predicted-versus-real CVR gap in the high-pCVR region, improving the prediction inputs on which OCPC’s bid boundaries depend.Values above the calibration threshold are calibrated; the typical threshold is t_c = 0.012.
4 MODEL ESTIMATION
Taobao’s OCPC relies on CTR and CVR prediction models built from sparse, high-dimensional user, context, and campaign features. The paper evaluates these models with GAUC alongside AUC and reports stable, practicable performance for production use.
- OCPC performance depends strongly on CVR prediction, while other predicted values such as pCTR also affect the strategy.
- 4.1 Model and Features: MLR is used in the RTP server because Taobao’s user and campaign features are sparse, high-dimensional, and potentially nonlinear.The feature space has tens of millions of dimensions.
- 4.1 Model and Features: CTR and CVR prediction combine context, user, and campaign features, including their cross-products; CVR additionally uses click-quality model outputs.
- 4.1 Model and Features: CTR and CVR models use clicked impressions and clicked-converted impressions as positive samples respectively, with daily retraining to reduce day-to-day variance.
- 4.2 Model Performance: GAUC evaluates weighted AUC across user–ad-position groups, and the reported MLR models show stable daily performance and practicable CTR/CVR prediction.Groups containing only positive or only negative samples are excluded; weights are proportional to impressions or clicks.
5 EXPERIMENTAL RESULTS
Offline simulation and online experiments evaluate OCPC against fixed and alternative bid strategies. OCPC improves platform, advertiser, and traffic-quality outcomes across overall, campaign, category, and scenario analyses.
- Offline simulation: Only Strategy 2 achieves simultaneous improvement in GPM, ROI, and RPM; Strategy 1 cannot ensure better RPM, while Strategy 3 lowers RPM.The comparison uses Strategy 0 as the non-optimized baseline and sets adjustment-function parameters through cross-validation.
- Offline simulation: Larger adjustment ranges increase performance in Strategy 2, with RPM increasing less than GPM and therefore producing a larger ROI lift.This finding comes from offline simulations varying the adjustment range ra.
- Campaign results: In campaign-level simulation, seven of ten campaigns increase GPM while reducing PV, and eight improve ROI under OCPC.The reduced PV indicates fewer low-quality opportunities, while campaigns 3 and 8 experience slight ROI declines because they compete for more PVs.
- Online results: In production traffic, Strategy 2 delivers a stable tripartite improvement in advertisers’ ROI, platform RPM, and overall GPM against Strategy 0.The online A/B test assigns 30% of production traffic to each strategy, with random user allocation and all campaigns present in both groups.
- Online results: Among campaigns with more than five weekly conversions, 67% improve both GPM and ROI, while category-level results show 17% of categories improving both metrics across 62% of PVs.A further production analysis finds traffic-share variations remain within ±10%, and OCPC also improves ASR by 15.6% in an alternative scenario.
6 CONCLUSION
The paper selects CPC for Taobao’s banner and item advertising formats and proposes OCPC to reconcile advertiser, platform, and ecological objectives. Production deployment and additional experiments support OCPC across traffic types and advertising scenarios.
- Conclusion: CPC is selected because it controls advertisers’ click costs while fitting Taobao’s data-rich, standardized e-commerce advertising process.The paper contrasts CPC with CPS, which ignores click value, and CPM, which poses higher risk for the mainly small and medium-sized advertisers.
- Conclusion: Additional production experiments evaluate OCPC in item ads, banner ads, and shopping-cart prediction settings.The supplied results include online experiments with 5% production flow for shopping-cart prediction and 30% production flow for banner ads.
- Conclusion: OCPC reconciles advertiser demands, platform ecological indices, and platform revenue while retaining eCPM sorting through bid optimization.It is automatically applied to mobile Item CPC Ads and can be selected for Banner CPC Ads.