Source-linked AI summary
Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate
Xiao Ma, Liqin Zhao, Guan Huang, Zhi Wang, Zelin Hu, Xiaoqiang Zhu, Kun Gai
TL;DR
CVR modeling must handle sample selection bias from training on clicked impressions while inferring over all impressions, alongside severe data sparsity. ESMM uses sequential impression–click–conversion modeling with auxiliary CTR and CTCVR tasks and feature transfer learning. On Taobao recommender-system data, ESMM consistently outperforms competitive models, and the authors release a sampled dataset for future research.
Problem
CVR models face sample selection bias from clicked-only training and data sparsity that makes fitting difficult.
Method
ESMM estimates pCTR and pCTCVR over the entire impression space, treats pCVR as an intermediate variable, and transfers feature representations from the CTR network.
Results
ESMM consistently outperforms competitive models on Taobao recommender-system datasets, achieving absolute AUC gains of 2.18% on CVR and 2.32% on CTCVR over BASE.
Takeaways & Limitations
Sequential action modeling with auxiliary CTR and CTCVR tasks provides an approach for CVR modeling that addresses sample selection bias and data sparsity simultaneously.
Takeaways & Limitations
The work does not focus on delayed feedback, although the authors state their approach can be combined with prior work to handle it.
Abstract
from arXiv · showhide
Estimating post-click conversion rate (CVR) accurately is crucial for ranking systems in industrial applications such as recommendation and advertising. Conventional CVR modeling applies popular deep learning methods and achieves state-of-the-art performance. However it encounters several task-specific problems in practice, making CVR modeling challenging. For example, conventional CVR models are trained with samples of clicked impressions while utilized to make inference on the entire space with samples of all impressions. This causes a sample selection bias problem. Besides, there exists an extreme data sparsity problem, making the model fitting rather difficult. In this paper, we model CVR in a brand-new perspective by making good use of sequential pattern of user actions, i.e., impression -> click -> conversion. The proposed Entire Space Multi-task Model (ESMM) can eliminate the two problems simultaneously by i) modeling CVR directly over the entire space, ii) employing a feature representation transfer learning strategy. Experiments on dataset gathered from Taobao's recommender system demonstrate that ESMM significantly outperforms competitive methods. We also release a sampling version of this dataset to enable future research. To the best of our knowledge, this is the first public dataset which contains samples with sequential dependence of click and conversion labels for CVR modeling.
1 INTRODUCTION
CVR prediction supports industrial ranking, but conventional models face sample selection bias and data sparsity because they train on clicked impressions while inferring across all impressions. ESMM uses sequential user actions and multi-task transfer learning to address both challenges, outperforming competitive models on Taobao data.
- CVR prediction helps industrial ranking systems balance users’ click and purchase preferences in advertising and recommendation.
- Post-click CVR estimates conversion probability conditioned on click and impression within the sequence impression → click → conversion.
- Conventional CVR models train on clicked impressions but infer across all impressions, creating sample selection bias.
- Existing approaches address sparsity or sampling sensitivity, but do not exploit the sequential dependence of user actions.
- ESMM introduces CTR and CTCVR auxiliary tasks, estimates both across the entire impression space, and derives pCVR as an intermediate variable.
- ESMM shares feature-representation parameters with the richer CTR task, using transfer learning to alleviate data sparsity while addressing sample selection bias.
- 8.9 billion Taobao recommender-system samples with sequential click and conversion labels support experiments in which ESMM consistently outperforms competitive models.
2 THE PROPOSED APPROACH
ESMM addresses sample selection bias and data sparsity in post-click CVR modeling by exploiting the sequential relationship among impressions, clicks, and conversions. It estimates associated tasks over the entire impression space while transferring feature representations from CTR modeling.
- CVR Modeling and Challenges: Post-click CVR is p(z = 1|y = 1,x), where clicks precede conversions in the user-action sequence.
- CVR Modeling and Challenges: Conventional CVR models train on clicked impressions but infer over all impressions, creating sample selection bias because the clicked-sample space may differ from the entire space.
- CVR Modeling and Challenges: CVR training data is extremely sparse because clicks are rare; the CVR sample count can be 1-3 orders of magnitude lower than CTR data and was 4% of CTR samples in the experiments.
- Entire Space Multi-Task Model: ESMM introduces CTR and CTCVR auxiliary tasks, sharing embeddings between CTR and CVR networks while producing CTCVR from their output product.
- Entire Space Multi-Task Model: ESMM models pCTR and pCTCVR using all impressions, treats pCVR as an intermediate variable, and uses their multiplication relationship to derive pCVR over the entire input space.
- Entire Space Multi-Task Model: Shared embedding parameters transfer representations learned from richer CTR data, allowing the CVR network to learn from un-clicked impressions and alleviate data sparsity.
3 EXPERIMENTS
Experiments compare ESMM with competitive CVR models on public and large-scale Taobao datasets across CVR and CTCVR tasks and sampling rates. ESMM consistently performs best, with substantial AUC gains over BASE.
- Datasets: The experiments use Taobao traffic data, including a 1% public sample and an 8.9-billion-sample Product Dataset.The released Public Dataset is 38GB without compression; the Product Dataset contains the full traffic logs.
- Competitors: Compared methods include BASE, AMAN, OVERSAMPLING, UNBIAS, DIVISION, ESMM-NS, and ESMM under shared network structures and hyperparameters.ESMM-NS removes embedding-parameter sharing from ESMM, while DIVISION separately trains CTR and CTCVR networks and divides their outputs.
- Evaluation: Models are evaluated on conventional CVR prediction with clicked impressions and CTCVR prediction over all impressions.Both tasks use the first half of the time sequence for training and the remainder for testing.
- Results on Public Dataset: 2.56% absolute AUC gain on CVR and 3.25% on CTCVR over BASE are achieved by ESMM on the Public Dataset.ESMM-NS improves over DIVISION by avoiding numerical instability, and ESMM further improves ESMM-NS through sequential-pattern modeling and transfer learning.
- Results on Product Dataset: ESMM-NS and ESMM consistently outperform competitors across sampling rates on the Product Dataset.All methods generally improve as training volume grows, indicating the influence of data sparsity.
4 CONCLUSIONS AND FUTURE WORK
The paper concludes that ESMM addresses sample selection bias and data sparsity through auxiliary CTR and CTCVR tasks, while suggesting extension to multistage user actions.
- Conclusions: ESMM uses sequential user-action patterns and auxiliary CTR and CTCVR tasks to address sample selection bias and data sparsity.Experiments on real data demonstrate superior performance, and the method can generalize to sequential user-action prediction.
- Future Work: Future work will design global optimization models for multistage actions such as request → impression → click → conversion.