# Learning Personalized Page Content Ranking Using Customer Representation

**Authors:** Xin Shen, Yan Zhao, Sujan Perera, Yujia Liu, Jinyun Yan, Mitchell Goodman  
**arXiv:** [2305.05267](https://arxiv.org/abs/2305.05267)  
**Date:** 2023-05-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.05267v1/

## Key points

- Authors Xin Shen, Yan Zhao, Sujan Perera, Yujia Liu, Jinyun Yan, and Mitchell Goodman explore challenges in providing relevant content recommendations on E-commerce platforms like Amazon and eBay.
- They introduce a content ranking service powered by a linear causal bandit algorithm that ranks and selects content for each shopper based on their individual context to address information overload.
- The algorithm relies on aggregated customer behavior features but recognizes the importance of inferring shoppers' interests from historical activities for better personalization.
- They propose a deep learning-based bandit algorithm incorporating historical shopping behavior, customer latent shopping goals, and correlation between customers and content categories.
- The model achieved a 12.08% lift in normalized discounted cumulative gain (nDCG) and demonstrated a 0.02% annualized commercial impact in online A/B testing scenarios.
- Experiments with categorical features showed improvements in nDCG@5 by +2.78% when incorporating personalized features compared to traditional approaches like deep neural networks or Deep Interest Network with traditional attention layers.
- The study highlights the significance of incorporating historical shopping behavior and customer preferences into recommendation systems for tailored shopping experiences.

## Summary

In the study "Learning Personalized Page Content Ranking Using Customer Representation," authors Xin Shen, Yan Zhao, Sujan Perera, Yujia Liu, Jinyun Yan, and Mitchell Goodman explore the challenges of providing relevant content recommendations on E-commerce platforms such as Amazon and eBay. The authors introduce a content ranking service powered by a linear causal bandit algorithm that ranks and selects content for each shopper based on their individual context. This approach addresses the issue of information overload for consumers due to the sheer volume of products available on these platforms. The algorithm primarily relies on aggregated customer behavior features rather than focusing on individual shopper-level past activities. However, the authors recognize the importance of inferring shoppers' interests from historical activities to improve the personalization of content recommendations. To achieve this goal, they propose a deep learning-based bandit algorithm that incorporates historical shopping behavior, customer latent shopping goals, and the correlation between customers and content categories. Through their model, the authors were able to achieve a 12.08% lift in normalized discounted cumulative gain (nDCG), indicating a significant improvement in personalized content ranking. In online A/B testing scenarios, the model demonstrated a 0.02% annualized commercial impact based on business metrics, validating its effectiveness in real-world applications. Additionally, the authors conducted experiments with categorical features using Mean Squared Error (MSE) loss reduction and Mean Absolute Error (MAE) reduction metrics. They observed improvements in nDCG@5 by +2.78% when incorporating personalized features into their model structure compared to traditional approaches like deep neural networks (NN) or Deep Interest Network with traditional attention layers (DIN-B). Overall,the study highlights  as an important aspect of , and the effectiveness of using a  and a . The proposed  with split-channel-wise attention emerged as a promising approach for improving both model fitting performance and personalized ranking accuracy. This research underscores the significance of incorporating historical shopping behavior and customer preferences into recommendation systems to provide more tailored shopping experiences for consumers.