Research map / Label-efficient and robust learning / Recognition and long-tail
Entire-Space Variational Information Exploitation for Post-Click Conversion Rate Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes EVI, an entire-space CVR framework that trains a click-conditioned CVR teacher for unbiased pseudo labels and maximizes variational mutual information plus logit distillation to transfer non-click information to the CVR student.
Problem
Post-click conversion rate (CVR) estimation in recommender systems suffers from sample selection bias (SSB) and data sparsity (DS); existing entire-space methods use biased pseudo labels for unclicked samples.
Approach
Condition the CVR teacher on click propensity (pCTR) to produce unbiased pseudo labels for unclicked samples, and transfer information at both logit and feature levels via a variational mutual-information lower bound (Gaussian variational distribution) plus logit distillation.
Key results
- EVI achieves AUC 0.6802 on Ali-CCP and 0.8292 on the industrial dataset, the best among all compared methods on all six datasets.
- EVI reports an average 2.25% AUC improvement and 2.78% NLL reduction over the best baselines.
- Ablation shows removing variational information exploitation and the conditional teacher each lowers AUC on all datasets (e.g., Ali-CCP 0.6702 without both vs 0.6802 full).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add variational information exploitation (VIE) (ablation) | EVI with CECT · Ali-CCP, AE-FR, AE-ES, AE-US, AE-NL, Industrial | AUC: 0.6776 -> 0.6802; 0.7971 -> 0.8008; 0.8075 -> 0.8140; 0.8015 -> 0.8057; 0.7717 -> 0.7741; 0.8254 -> 0.8292 | Helped |
| add conditional entire-space CVR teacher (CECT) (ablation) | EVI without VIE · Ali-CCP, AE-FR, AE-ES, AE-US, AE-NL, Industrial | AUC: 0.6702 -> 0.6776; 0.7777 -> 0.7971; 0.7943 -> 0.8075; 0.7888 -> 0.8015; 0.7675 -> 0.7717; 0.8218 -> 0.8254 | Helped |
| use EVI rather than the strongest baseline | CVR prediction · Ali-CCP, AE-FR, AE-ES, AE-US, AE-NL, Industrial | AUC | Helped |
| use EVI rather than the strongest baseline | CVR prediction · Ali-CCP, AE-FR, AE-ES, AE-US, AE-NL, Industrial | NLL | Helped |
Surprising findings
- The paper reports that DDPO shows less mean bias than UKD on Ali-CCP and AE-FR despite having higher pseudo-label logloss, attributed to its IPS weighting.
Limitations
- EVI relies on accurate CTR estimation, and its unbiasedness theorems depend on accurate click propensities and pseudo labels.
- The method is sensitive to noisy clicked samples.
- Reported improvements rely on hyperparameter search over trade-off weights using test-set average AUC.
Experimental setup
Datasets: Ali-CCP AE-FR AE-ES AE-US AE-NL Industrial affiliate advertising dataset
Benchmarks: Ali-CCP AE-FR AE-ES AE-US AE-NL Industrial
Metrics: AUC NLL
Baselines: ESMM ESCM2-IPW ESCM2-DR MRDR-GPL DR-V2 UKD DCMT DDPO
Builds on: ESMM UKD DDPO ESCM2 MMOE
Closest papers by meaning
- Online Conversion Rate Prediction via Multi-Interval Screening and Synthesizing under Delayed Feedback (AAAI 2024 · similarity 0.85)
- AT4CTR: Auxiliary Match Tasks for Enhancing Click-Through Rate Prediction (AAAI 2024 · similarity 0.84)
- Addressing Cold-Start Problem in Click-Through Rate Prediction via Supervised Diffusion Modeling (AAAI 2025 · similarity 0.83)
- Addressing Hidden Confounding with Heterogeneous Observational Datasets for Recommendation (NeurIPS 2024 · similarity 0.82)
- Fine-Grained Dynamic Framework for Bias-Variance Joint Optimization on Data Missing Not at Random (NeurIPS 2024 · similarity 0.82)
- Towards Unbiased Information Extraction and Adaptation in Cross-Domain Recommendation (AAAI 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.