# Estimating the Causal Impact of Recommendation Systems from Observational Data

**Authors:** Amit Sharma, Jake M. Hofman, Duncan J. Watts  
**arXiv:** [1510.05569](https://arxiv.org/abs/1510.05569)  
**Date:** 2015-10-19  
**Summary page:** https://www.summarizepaper.com/en/paper/1510.05569v1/

## Key points

- Recommendation systems account for a significant portion of web traffic on popular websites
- The paper proposes a method for estimating the causal impact of recommendation systems using observational data
- The authors demonstrate that causal identification is possible when direct traffic experiences an instantaneous shock while recommended products do not
- The method is applied to browsing logs from 2.1 million users on Amazon.com over a 9-month period
- Findings show that at least 75% of recommendation click-throughs would likely occur even without recommendations
- The study cautions against extrapolating the results to other products, sites, or settings without considering limitations

## Summary

The paper titled "Estimating the Causal Impact of Recommendation Systems from Observational Data" by Amit Sharma, Jake M. Hofman, and Duncan J. Watts explores the impact of recommendation systems on web traffic. <br>
Recommendation systems have become increasingly prominent on popular websites, accounting for a significant portion of their overall traffic. However, it is unclear how much additional activity these systems generate compared to what would occur through other means such as search.<br>
While randomized experiments are the ideal method to estimate the causal impact of recommendations, they can be expensive and inconvenient for users. In this paper, the authors propose a method for estimating causal effects using purely observational data. They demonstrate that causal identification through an instrumental variable is possible when a product experiences an instantaneous shock in direct traffic while the recommended products do not.<br>
To analyze the impact of recommendation systems, the authors apply their method to browsing logs from 2.1 million users on Amazon.com over a 9-month period. They focus on over 4,000 unique products that experience sudden increases in direct traffic but no corresponding increase in recommended products.<br>
The findings reveal that although recommendation click-throughs contribute significantly to traffic among these products, at least 75% of this activity would likely occur even without recommendations.<br>
The paper concludes with a discussion about the assumptions underlying their method and cautions against extrapolating the results to other products, sites, or settings without considering potential limitations.<br>
Overall, this study provides valuable insights into understanding the causal impact of recommendation systems on web traffic using observational data. It highlights that while recommendations play a substantial role in driving user activity, a significant portion of this activity would still occur independently of recommendations.