# A Practical Method for Solving Contextual Bandit Problems Using Decision Trees

**Authors:** Adam N. Elmachtoub, Ryan McNellis, Sechan Oh, Marek Petrik  
**arXiv:** [1706.04687](https://arxiv.org/abs/1706.04687)  
**Date:** 2017-06-14  
**Summary page:** https://www.summarizepaper.com/en/paper/1706.04687v1/

## Key points

- The contextual multi-armed bandit problem has efficient algorithms with strong theoretical guarantees
- Applying these algorithms in practice is challenging due to the need for domain expertise to build appropriate features and tune their parameters
- Adam N. Elmachtoub, Ryan McNellis, Sechan Oh, and Marek Petrik proposed a new method that is simple and practical
- This method relies on decision trees to model the context-reward relationship as they are non-parametric and interpretable
- A bootstrapping approach is used to guide the exploration-exploitation trade-off which abstracts Thompson sampling to non-Bayesian settings
- The algorithm outperforms existing methods while being simpler to implement and requiring less domain expertise
- This paper presents a practical solution for solving contextual bandit problems using decision trees that can be applied with ease by practitioners without extensive knowledge of machine learning or statistics.

## Summary

The contextual multi-armed bandit problem has been the subject of many efficient algorithms with strong theoretical guarantees. To address the challenge of applying these algorithms in practice due to the need for domain expertise to build appropriate features and tune their parameters, Adam N. Elmachtoub, Ryan McNellis, Sechan Oh, and Marek Petrik proposed a new method that is simple and practical. This method relies on decision trees to model the context-reward relationship as they are non-parametric and interpretable. A bootstrapping approach is used to guide the exploration-exploitation trade-off which abstracts Thompson sampling to non-Bayesian settings. The authors discuss several computational heuristics in detail in the paper which demonstrate how their algorithm performs on several datasets. It was found that this algorithm outperforms existing methods while being simpler to implement and requiring less domain expertise. Overall, this paper presents a practical solution for solving contextual bandit problems using decision trees that can be applied with ease by practitioners without extensive knowledge of machine learning or statistics. This approach provides a promising avenue for future research in this area as it offers a more accessible way of tackling contextual bandit problems while maintaining high levels of performance.