# Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate

**Authors:** Xiao Ma, Liqin Zhao, Guan Huang, Zhi Wang, Zelin Hu, Xiaoqiang Zhu, Kun Gai  
**arXiv:** [1804.07931](https://arxiv.org/abs/1804.07931)  
**Date:** 2018-04-21  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/1804.07931v1/

## Key points

- The paper addresses the challenge of accurately estimating post-click conversion rate (CVR) in industrial applications such as recommendation and advertising.
- Conventional CVR modeling techniques face challenges including sample selection bias and extreme data sparsity problems.
- The authors propose a new approach called Entire Space Multi-task Model (ESMM) that leverages the sequential pattern of user actions: impression -> click -> conversion.
- ESMM directly models CVR over the entire space and employs a feature representation transfer learning strategy to eliminate sample selection bias and extreme data sparsity problems.
- Experiments on a dataset from Taobao's recommender system show that ESMM outperforms competitive methods significantly.
- The authors release a sampling version of the dataset for future research.

## Summary

The paper titled "Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate" addresses the challenge of accurately estimating post-click conversion rate (CVR) in industrial applications such as recommendation and advertising. While conventional CVR modeling techniques based on deep learning methods have achieved state-of-the-art performance, they face several task-specific problems in practice. One of the main challenges is the sample selection bias problem. Conventional CVR models are trained using samples of clicked impressions but are then used to make inferences on the entire space with samples of all impressions. This mismatch between training and inference data causes a bias in the model's performance. Another challenge is the extreme data sparsity problem. The scarcity of data makes it difficult to fit the CVR model effectively, leading to suboptimal performance. To address these challenges, the authors propose a brand-new perspective for modeling CVR by leveraging the sequential pattern of user actions: impression -> click -> conversion. They introduce the Entire Space Multi-task Model (ESMM), which directly models CVR over the entire space and employs a feature representation transfer learning strategy. The ESMM approach simultaneously eliminates both the sample selection bias problem and the extreme data sparsity problem. By modeling CVR directly over the entire space, ESMM ensures that inferences are made using samples representative of all impressions. Additionally, employing a feature representation transfer learning strategy helps improve model fitting even with limited data. Experiments conducted on a dataset gathered from Taobao's recommender system demonstrate that ESMM outperforms competitive methods significantly. The authors also release a sampling version of this dataset to facilitate future research. Overall, this paper presents an innovative approach to CVR modeling that tackles key challenges faced by conventional methods and provides valuable insights into improving ranking systems in industrial applications such as recommendation and advertising.