Abstract
Estimating post-click conversion rate (CVR) is a critical task in industrial recommendation systems. Traditional CVR modeling approaches face two major challenges: Sample Selection Bias (SSB) and Data Sparsity (DS). This paper introduces the Entire Space Multi-Task Model (ESMM), which addresses these issues by modeling CVR over the entire impression space. By leveraging the sequential nature of user actions (Impression → Click → Conversion) and employing multi-task learning with shared feature representations, ESMM effectively eliminates SSB and mitigates the impact of DS.
Introduction: The Challenges of CVR Estimation
In e-commerce, the user journey typically follows a sequence: an item is shown (Impression), the user may interact with it (Click), and finally, they may purchase it (Conversion). Conventional CVR models are typically trained only on samples that have been clicked. However, during the inference stage, the model is required to rank all items presented in the impression space.
Sample Selection Bias and Data Sparsity
This discrepancy between training and inference leads to Sample Selection Bias (SSB), where the training distribution significantly differs from the real-world inference distribution. Furthermore, because conversions are much rarer than clicks (often by a factor of 10 to 100), models suffer from extreme Data Sparsity (DS), making it difficult to learn complex deep learning representations without overfitting.
Methodology: The Entire Space Multi-Task Model (ESMM)
To address these issues, the ESMM framework leverages the mathematical relationship between three probabilities:
- pCTR: Probability of a click given an impression.
- pCVR: Probability of a conversion given a click.
- pCTCVR: Probability of both a click and a conversion given an impression.
The core insight is the multiplicative formula: $pCTCVR = pCTR \times pCVR$.
Model Architecture
ESMM consists of two sub-networks: a CTR network and a CVR network. Instead of training the CVR network in isolation on clicked data, ESMM optimizes for CTR and CTCVR tasks simultaneously over the entire impression space. The pCVR is treated as an intermediate variable derived such that the product of the two networks' outputs matches the CTCVR label.
A key innovation of ESMM is the Shared Lookup Table. The embedding layer is shared between the CTR and CVR tasks. Since the CTR task has access to significantly more data (all impressions), it learns robust feature representations for users and items that the data-starved CVR task can "borrow," effectively addressing the data sparsity problem through transfer learning.
Experimental Results
The researchers evaluated ESMM using two massive datasets from Taobao: an internal production dataset containing 8.9 billion samples and a public benchmark dataset.
Performance and Robustness
On the public dataset, ESMM achieved an absolute AUC gain of 2.56% on the CVR task compared to standard baseline models. In industrial settings, even gains of 0.1% are considered significant. Furthermore, the model demonstrated remarkable robustness to data sparsity; as the training set size decreased, ESMM's performance remained stable, significantly outperforming alternative methods like oversampling or unbiasing techniques.
Conclusion
The Entire Space Multi-Task Model (ESMM) provides a robust solution for CVR estimation by aligning the training process with the physical reality of user behavior. By modeling over the entire impression space and sharing feature representations, ESMM overcomes the fundamental hurdles of sample selection bias and data sparsity. This approach sets a new benchmark for sequential behavior modeling in large-scale recommendation systems.
Summary updated