Entire-Space Variational Information Exploitation for Post-Click Conversion Rate Prediction

Ke Fei, Xinyue Zhang, Jingjing Li

Abstract

Predicting Post-Click Conversion Rate (CVR) is critical for recommendation systems but faces challenges from sample selection bias (SSB) and data sparsity (DS). This paper introduces the Entire-Space Variational Information Exploitation (EVI) framework. EVI addresses SSB by training an unbiased "teacher" model conditioned on click propensity and mitigates DS by transferring rich feature-level knowledge to a "student" model through variational information maximization. Experimental results across multiple large-scale datasets show that EVI significantly outperforms existing state-of-the-art methods, achieving an average AUC improvement of 2.25%.

Introduction

In modern recommendation systems, predicting the Post-Click Conversion Rate (CVR)—the probability of a purchase following a click—is essential for optimizing business value. However, two major obstacles persist: Sample Selection Bias (SSB) and Data Sparsity (DS). SSB occurs because models are typically trained only on clicked items, yet must perform across the entire impression space. DS arises because conversions are far rarer than clicks, providing limited signal for learning. While "entire-space" methods attempt to use unclicked samples via pseudo-labels, these labels often inherit biases by ignoring the underlying click propensity. The EVI framework tackles these issues by ensuring unbiased knowledge transfer from a teacher to a student model.

Methodology

The proposed EVI framework introduces a teacher-student distillation pipeline with two core innovations: the Conditional Entire-Space CVR Teacher (CECT) and Variational Information Exploitation (VIE).

The EVI framework architecture, highlighting the Conditional Entire-Space CVR Teacher (CVR-T) and the Variational Information Maximization (VIE) mechanism.

The CECT leverages causal inference principles, specifically "do-calculus," to condition the teacher model on click propensity. This ensures the teacher's CVR estimates are mathematically unbiased across the entire impression space. To transfer this knowledge effectively, VIE goes beyond standard logit distillation. It employs a variational information maximization scheme to ensure the student model's internal layers retain maximal information from the teacher's representations, allowing the student to exploit patterns learned from the non-click space. The final model is optimized using a multi-task loss function that balances Click-Through Rate (CTR), CVR, and the variational information transfer.

Experimental Results

EVI was evaluated on six large-scale datasets, including Ali-CCP, Ali-Express (AE), and an Industrial dataset from a major advertising system. These datasets are characterized by their sequential nature (impression → click → conversion) and extreme sparsity. The framework was compared against eight competitive baselines, such as ESMM and UKD.

Comparison of Logloss and Mean Bias across Ali-CCP, AE-FR, and Industrial datasets, demonstrating EVI's superior performance and reduced bias.

Key Findings

  • Performance Gains: EVI achieved an average AUC improvement of 2.25% and a 2.78% reduction in Logloss across all datasets.
  • Bias Reduction: Experimental results confirmed that EVI’s teacher produces significantly lower Logloss and Mean Bias on unclicked samples compared to other state-of-the-art teachers like UKD or DDPO. This validates the theoretical claim that conditioning on click propensity is key to unbiasedness.
  • Ablation Studies: Removing either the CECT or VIE components led to a marked decrease in performance, confirming that both unbiased conditioning and deep feature-level information transfer are critical.

Conclusion and Future Work

EVI represents a significant advancement in CVR prediction by theoretically addressing bias transfer. By maximizing mutual information between the teacher and student's hidden layers, the student learns the underlying logic of the teacher rather than just its final predictions. While the framework's effectiveness depends on the accuracy of the underlying CTR model and introduces some training-time complexity, its ability to learn from the entire impression space provides a robust solution to the sparsity and bias problems. Future research will explore more robust ways to handle "noisy" clicks—instances where a user clicks by accident—which can further distort CVR estimation.

Summary updated

BibTeX
@inproceedings{fei2025entire,
  author = {Ke Fei and Xinyue Zhang and Jingjing Li},
  title = {{Entire-Space Variational Information Exploitation for Post-Click Conversion Rate Prediction}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i11.33268},
}
Figure