# Off-policy evaluation for slate recommendation

**Authors:** Adith Swaminathan, Akshay Krishnamurthy, Alekh Agarwal, Miroslav Dudík, John Langford, Damien Jose, Imed Zitouni  
**arXiv:** [1605.04812](https://arxiv.org/abs/1605.04812)  
**Date:** 2016-05-16  
**Summary page:** https://www.summarizepaper.com/en/paper/1605.04812v1/

## Key points

- The paper focuses on off-policy evaluation for slate recommendation in contexts such as web search, ads, and recommender systems.
- The authors propose a novel technique for assessing policies offline using historical data with minimal bias.
- The approach is based on breaking down the overall quality of a recommended set into additive components across individual items.
- Despite limitations in directly observing or modeling the quality of each item based on its features, this assumption holds true in many practical scenarios.
- The authors demonstrate through theoretical analysis that their method enables exponential savings in data requirements compared to naive unbiased methods.
- By effectively leveraging past data and addressing complexities in evaluating recommendation policies, this research contributes valuable insights to the field of recommendation systems.

## Summary

The paper "Off-policy evaluation for slate recommendation" by Adith Swaminathan, Akshay Krishnamurthy, Alekh Agarwal, Miroslav Dudík, John Langford, Damien Jose, and Imed Zitouni delves into the evaluation of policies that recommend an ordered set of items in various contexts such as web search, ads, and recommender systems. The authors propose a novel technique for assessing these policies offline using historical data with minimal bias. Their approach is based on the assumption that the overall quality of a recommended set can be broken down into additive components across individual items. This assumption holds true in many practical scenarios despite the limitation that the quality of each item cannot be directly observed or modeled based on its features. Furthermore, through theoretical analysis, the authors demonstrate that their method enables exponential savings in data requirements compared to naive unbiased methods. By effectively leveraging past data and accounting for the complexities of evaluating recommendation policies, this research contributes valuable insights to the field of recommendation systems.