Practical Multi-Task Learning for Rare Conversions in Ad Tech
Proposes a data-driven multi-task learning (MTL) framework that enhances rare conversion prediction (<1%) in online advertising, achieving 0.69% AUC lift offline and 2% CPA reduction online.
Key Findings
Methodology
This paper introduces a deep neural network-based multi-task learning architecture featuring shared embedding and multiple task-specific towers. Tasks are automatically defined based on historical conversion rates (CVR), with a threshold (α) separating 'rare' and 'frequent' conversions. The model employs a weighted loss function (e.g., W_hard and W_soft) to balance learning between tasks, with shared layers capturing common representations and task-specific towers optimizing for each conversion type. During training, each sample contributes only to its assigned task's loss, facilitating robust learning amidst data imbalance. For inference, only the relevant task branch is retained, ensuring efficiency. The task classification leverages real-time CVR estimates and contextual features, routing ad candidates accordingly. To address stability issues caused by CVR fluctuations, a multi-dimensional smoothing mechanism is integrated. Extensive offline validation on hundreds of millions of logs shows a 0.69% AUC improvement, while online A/B testing demonstrates a 2% reduction in CPA and a 3% increase in conversions, validating the approach's effectiveness in production.
Key Results
- Offline evaluation on large-scale logs revealed a 0.69% AUC lift and 4.08% relative information gain (RIG) when using the optimized multi-task model compared to the baseline. Hyperparameter tuning identified W_hard=W_soft=0.5 as optimal, with more shared layers (up to 4) further enhancing performance. The model effectively leverages soft conversions as auxiliary signals, leading to better generalization on hard conversions. Online A/B tests confirmed these gains, with the multi-task model reducing CPA by 3.4% and increasing conversions by 3%, especially among high-spending advertisers, demonstrating significant business impact.
- The model's stability was validated through multi-dimensional smoothing and contextual adjustments, ensuring consistent performance despite CVR fluctuations. The automatic task definition based on historical CVR proved robust across different ad scenarios, reducing reliance on manual labels. The deployment included optimization for low signal scenarios, maintaining prediction accuracy and serving stability at scale.
Significance
This work addresses a long-standing challenge in online advertising: accurately predicting rare but valuable conversions. By leveraging multi-task learning with data-driven task definitions, the approach significantly improves prediction stability and accuracy for low-frequency events, directly enhancing ad campaign ROI. The automatic task segmentation reduces dependence on manual tagging, making the system scalable and adaptable to dynamic advertising environments. The successful large-scale deployment demonstrates practical viability, setting a new standard for rare event prediction in computational advertising. The methodology's generality suggests potential applications in other domains involving infrequent but impactful events, such as fraud detection or high-value user actions.
Technical Contribution
The paper's main technical innovation lies in integrating a data-driven, dynamic task definition mechanism within a deep multi-task learning framework. Unlike prior methods relying on manual labels, this approach uses historical CVR thresholds to automatically classify conversion signals, reducing label noise. The architecture combines shared embedding layers with multiple task-specific towers, enabling effective feature sharing and specialization. The loss function balances contributions from soft and hard tasks, tuned via hyperparameters. The model incorporates contextual smoothing to enhance stability, especially critical for low-signal scenarios. Extensive offline hyperparameter tuning and online deployment validate the approach's robustness and scalability. This work advances the state-of-the-art in rare event prediction by demonstrating how automatic task segmentation and multi-task learning can be combined to address data imbalance and stability issues in large-scale systems.
Novelty
This research is pioneering in applying a fully data-driven, automatic task definition for multi-task learning in the context of rare conversion prediction. Unlike previous works that manually label conversion types, this method dynamically classifies signals based on empirical CVR thresholds, reducing label noise and improving robustness. The architecture's integration of shared layers with task-specific towers, combined with contextual smoothing for stability, represents a novel engineering solution tailored for large-scale ad platforms. The successful deployment and online validation further distinguish this work, showcasing its practical impact and setting a new benchmark for future research in rare event modeling.
Limitations
- The approach relies heavily on the stability of historical CVR estimates; sudden market shifts or campaign changes can lead to misclassification of tasks, affecting prediction accuracy.
- In extremely sparse data scenarios, especially for ultra-rare events, the model may still face challenges due to insufficient signal, leading to potential instability or bias.
- The training and inference processes are computationally intensive, requiring significant resources for hyperparameter tuning and large-scale deployment, which may limit real-time adaptability in resource-constrained environments.
Future Work
Future research will explore adaptive, real-time task redefinition mechanisms that respond to rapid environment changes. Incorporating causal inference techniques could further mitigate biases arising from confounding factors. Additionally, efforts will focus on model compression and acceleration methods to reduce computational costs, enabling deployment in edge or latency-sensitive applications. Extending the framework to other rare event domains, such as fraud detection or high-value user actions, will be pursued. Lastly, integrating reinforcement learning for dynamic policy adjustment based on live feedback could further enhance prediction robustness and business outcomes.
AI Executive Summary
In the rapidly evolving landscape of digital advertising, accurately predicting user conversions remains a critical challenge. While frequent conversions such as page views or clicks are relatively easy to model due to abundant data, rare high-value events—like luxury purchases or subscription sign-ups—pose significant difficulties. These events occur at rates below 1%, resulting in severe data sparsity that hampers traditional predictive models, leading to unstable and biased estimates.
Existing solutions often rely on manual tagging or heuristic rules to distinguish between common and rare conversions. However, such approaches are prone to label noise and lack scalability, especially as advertising platforms handle billions of requests daily. Moreover, models trained without considering the inherent imbalance tend to underperform on rare events, limiting their utility for high-stakes advertisers.
To address these issues, the authors propose a novel multi-task learning (MTL) framework that leverages large-scale historical data to automatically define tasks based on empirical conversion rates. Instead of relying on manual labels, the system computes a threshold (α) on historical CVR to dynamically classify conversion signals into 'rare' and 'frequent' categories. This data-driven task segmentation allows the model to focus on improving predictions for the most valuable but infrequent conversions.
The core architecture features shared embedding layers and multiple task-specific towers, enabling the model to learn common representations while optimizing for each conversion type. During training, a weighted loss function balances the contributions of soft and hard tasks, with hyperparameters tuned via offline validation. In deployment, the model employs a routing mechanism that directs ad candidates to the appropriate task branch based on real-time CVR estimates, ensuring efficiency.
Extensive experiments on hundreds of millions of logs demonstrate the effectiveness of this approach. Offline metrics show a 0.69% increase in AUC and a 4.08% improvement in relative information gain, validating the model's enhanced predictive power. Online A/B testing further confirms its practical benefits, with a 2% reduction in CPA and a 3% increase in conversions, especially among high-value advertisers.
Beyond performance gains, the study addresses deployment challenges such as service stability and low signal robustness. By incorporating contextual smoothing and adaptive task classification, the system maintains consistent predictions even amidst CVR fluctuations. The successful large-scale deployment underscores the approach’s scalability and industry relevance.
Looking ahead, the authors plan to refine dynamic task definitions, incorporate causal inference, and explore model compression techniques. These advancements aim to further improve prediction accuracy, reduce computational costs, and extend applicability to other domains involving rare but impactful events. Overall, this work sets a new benchmark for rare event prediction in ad tech, with broad implications for digital marketing and beyond.
Deep Analysis
Background
The evolution of online advertising has seen a shift from simple click-based models to complex deep learning architectures capable of capturing intricate user-behavior patterns. Early models like logistic regression laid the groundwork, but their limited capacity to model feature interactions prompted the adoption of neural networks such as Deep & Cross Network (DCN). CVR prediction, crucial for post-click monetization, faces unique challenges due to data imbalance—high-frequency signals like landing page views contrast sharply with rare high-value conversions like luxury product purchases. Prior works, including ESMM and DCMT, attempted to jointly model click and conversion signals, reducing bias and improving stability. However, these models still struggle with the extreme sparsity and noise inherent in rare event data. Manual tagging of conversion types, often used to address heterogeneity, introduces label noise and scalability issues. Recent trends favor data-driven, automatic segmentation based on empirical conversion rates, enabling models to adapt dynamically to changing environments. Despite these advances, accurately predicting rare conversions remains a significant challenge, especially in large-scale, real-time systems where stability and efficiency are paramount.
Core Problem
The core problem addressed in this paper is the accurate prediction of rare but valuable conversion events (<1% CVR) in online advertising systems. Traditional models often underperform in this regime due to severe data imbalance, leading to unstable estimates and poor generalization. Manual tagging of conversion types is unreliable and labor-intensive, often resulting in noisy labels that degrade model performance. Furthermore, existing multi-task learning approaches typically rely on predefined, manually labeled tasks, limiting scalability and robustness. The challenge is to develop a scalable, automatic method that leverages large-scale historical data to dynamically define tasks, improve prediction accuracy for rare events, and ensure deployment stability in a real-time environment. Addressing these issues is critical for maximizing ROI for high-value advertisers and maintaining platform competitiveness in a data-saturated ecosystem.
Innovation
The paper introduces several key innovations:
1) Data-driven task definition: Instead of manual labels, the model uses historical CVR thresholds (α) to automatically classify conversion signals into 'rare' and 'frequent' categories, reducing label noise and improving robustness.
2) Adaptive multi-task architecture: Incorporates shared embedding layers and multiple task-specific towers, enabling effective feature sharing and specialization, which enhances the prediction of low-frequency events.
3) Dynamic routing and stability mechanisms: Utilizes real-time CVR estimates and contextual features to route ad candidates to appropriate task branches, coupled with smoothing techniques to address CVR fluctuations, ensuring prediction stability.
4) Large-scale deployment: Validated through extensive offline experiments and online A/B testing, demonstrating real-world effectiveness and scalability.
5) Optimization for low signal scenarios: Focuses on improving model robustness in environments with sparse data, crucial for rare event prediction in ad tech.
Methodology
- �� 数据准备：收集广告平台的海量日志数据，计算每个广告的历史CVR（采用时间衰减机制），作为任务划分的依据。• 任务定义：设定CVR阈值（α），低于该值的转化事件归为‘稀有’（hard），高于的为‘频繁’（soft），并根据实时CVR动态调整阈值以适应环境变化。• 模型架构：构建深度神经网络，包含共享嵌入层和多层全连接层，最后分为两个任务塔（soft和hard），每个塔专注于对应类别的预测。• 损失函数：采用加权多任务损失（L = W_soft·L_soft + W_hard·L_hard），超参数W_hard和W_soft通过离线验证调优。• 训练策略：使用mini-batch随机梯度下降，每个样本只贡献对应任务的损失，确保模型在不同类别间平衡学习。• 线上推理：在实际服务中，只保留目标任务分支，减少推理成本。• 任务分类：在广告请求到达时，根据实时CVR估算值和上下文信息，将广告路由到对应的任务模型，确保预测的针对性和效率。• 稳定性调节：引入多维平滑机制，结合广告类型、时间窗口等信息，缓解CVR波动带来的不稳定性。• 超参数调优：通过离线验证和线上AB测试，持续优化模型性能，确保在实际环境中的稳定性和效果。
Experiments
实验采用来自广告平台的数十亿条日志，划分为训练集和验证集，确保样本的多样性和代表性。对比基线模型（单任务、全量训练）和多任务模型（不同超参数配置），指标包括AUC、相对信息增益（RIG）以及线上关键指标（CPA、转化量）。超参数调优采用网格搜索，重点调节W_hard、W_soft和共享层数。离线验证显示，优化后模型在硬转化类别上实现了0.69%的AUC提升（从0.75到0.755），RIG提升4.08%。线上A/B测试在多个广告场景中验证了模型的实用性，平均降低CPA成本2%，转化量提升3%。还进行了边界样本的稳定性分析，确保模型在低信号环境下表现一致。模型训练过程中还进行了特征重要性分析和消融实验，验证了任务定义和共享层设计的有效性。
Results
模型在离线验证中表现优异，AUC从0.75提升到0.755，提升幅度达0.69%，验证了多任务学习在稀有事件预测中的有效性。相对信息增益（RIG）提升4.08%，说明模型在信息利用效率上有显著改善。超参数调优显示，W_hard和W_soft均衡设置（如0.5）效果最佳，充分利用软信号辅助硬转化预测。增加共享层数（最多4层）进一步提升模型性能，验证了共享表示的优势。线上AB测试中，采用多任务模型后，硬转化类别的CPA成本平均降低了3.4%，转化量增加了3%，特别是在高投入广告客户中效果更为显著。这些结果表明，模型不仅在指标上优于传统单任务模型，也在实际商业应用中带来了明显的收益。
Applications
该模型适用于广告平台中的稀有高价值转化事件预测，能帮助广告主优化投放策略，降低成本，提高ROI。模型部署后，能实时处理海量广告请求，提升广告排名和出价效率。未来还可结合多模态特征和实时反馈，进一步增强模型的适应性和预测能力。该技术也可推广到其他行业，如电商中的高端商品购买预测、金融中的异常交易检测等，具有广泛的应用潜力。模型还可以作为广告系统中的核心组件，支持个性化推荐和预算优化，为广告主提供更精准的投放建议，推动行业数字化转型。
Limitations & Outlook
模型高度依赖于历史CVR的稳定性，若广告环境发生剧烈变化（如市场波动、广告策略调整），任务划分可能失效，影响预测效果。极端低频场景中，数据稀疏问题仍然存在，模型可能在某些特殊场景下表现不佳。训练和推理过程的计算成本较高，尤其在超大规模数据环境中，需投入大量资源进行超参数调优和模型压缩。此外，模型对新兴广告类型和行业场景的泛化能力有限，未来需要引入迁移学习和在线学习机制，以增强模型的适应性和鲁棒性。最后，模型在极端偏差环境下可能出现过拟合或偏差，需要持续监控和调整。
Plain Language Accessible to non-experts
想象你在一家大型工厂工作，工厂每天生产各种商品。有些商品很常见，比如普通的玩具，每天都能卖出去很多；而有些商品很稀少，比如定制的奢侈品，卖得很少但利润很高。工厂的目标是预测哪些商品会卖得好，从而合理安排生产和销售。传统的方法就像用一个简单的统计表，只看过去卖了多少，但这样很难准确预测稀有商品的销售，因为它们太少，数据不够充分。
于是，工厂引入了一个聪明的系统，它不仅看稀有商品的销售，还会参考那些常见商品的销售情况，把两者结合起来。这个系统会自动判断某个商品是“稀有”还是“常见”，然后用不同的方法预测它们的销售。它还会根据过去的销售数据不断调整自己的判断，确保预测更准确。
这个系统的核心思想就像是一个聪明的助手，能从大量的销售记录中学习，知道哪些商品可能会卖得少但利润高，哪些商品卖得多。这样，工厂就能更好地准备库存，避免浪费，也能抓住那些稀有但高价值的机会。这个方法不仅让预测变得更聪明，还能在实际生产和销售中带来更大的收益。
Abstract
We present a Multi-Task Learning (MTL) approach for improving predictions for rare (e.g., <1%) conversion events in online advertising. The conversions are classified into "rare" or "frequent" types based on historical statistics. The model learns shared representations across all signals while specializing through separate task towers for each type. The approach was tested and fully deployed to production, demonstrating consistent improvements in both offline (0.69% AUC lift) and online KPI performance metric (2% Cost per Action reduction).
References (16)
Click-Conversion Multi-Task Model with Position Bias Mitigation for Sponsored Search in eCommerce
Yibo Wang, Yanbing Xue, Boyang Liu et al.
DCMT: A Direct Entire-Space Causal Multi-Task Framework for Post-Click Conversion Estimation
Feng Zhu, Mingjie Zhong, Xinxing Yang et al.
ESCM2: Entire Space Counterfactual Multi-Task Model for Post-Click Conversion Rate Estimation
Hao Wang, Tai-Wei Chang, Tianqiao Liu et al.
DCN V2: Improved Deep & Cross Network and Practical Lessons for Web-scale Learning to Rank Systems
Ruoxi Wang, Rakesh Shivanna, D. Cheng et al.
Predicting Different Types of Conversions with Multi-Task Learning in Online Advertising
Junwei Pan, Yizhi Mao, A. L. Ruiz et al.
Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate
Xiao Ma, Liqin Zhao, Guan Huang et al.
A Practical Framework of Conversion Rate Prediction for Online Display Advertising
Quan Lu, Shengjun Pan, Liang Wang et al.
Deep & Cross Network for Ad Click Predictions
Ruoxi Wang, Bin Fu, G. Fu et al.
An Overview of Multi-Task Learning in Deep Neural Networks
Sebastian Ruder
TensorFlow: A system for large-scale machine learning
Martín Abadi, P. Barham, Jianmin Chen et al.
Simple and Scalable Response Prediction for Display Advertising
O. Chapelle, Eren Manavoglu, Rómer Rosales
Predictive model performance: offline and online evaluations
Jeonghee Yi, Ye Chen, Jie Li et al.
Estimating conversion rate in display advertising from past erformance data
Kuang-chih Lee, Burkay Orten, A. Dasdan et al.
Estimating rates of rare events with multiple hierarchies through scalable log-linear models
Deepak Agarwal, Rahul Agrawal, Rajiv Khanna et al.
Multitask Learning
R. Caruana
The meaning and use of the area under a receiver operating characteristic (ROC) curve.
James A. Hanley, Barbara J. McNeil
Cited By (3)
Multi-Objective Ranking for Live-Streaming: Balancing Fresh and Delayed Signals with Segment-Aware Targeting
RAMP: Robust Ad Recommendation Under Limited Personalized-Feature Availability via Masking and Alignment Pathways
Conversion rate prediction in online advertising: modeling techniques, performance evaluation and future directions