# Post-Estimation Smoothing: A Simple Baseline for Learning with Side Information

**Authors:** Esther Rolf, Michael I. Jordan, Benjamin Recht  
**arXiv:** [2003.05955](https://arxiv.org/abs/2003.05955)  
**Date:** 2020-03-12  
**Summary page:** https://www.summarizepaper.com/en/paper/2003.05955v1/

## Key points

- Authors explore significance of natural structural indices in observational data
- Proposal of a novel approach to leverage structural index data in prediction tasks
- Introduction of post-estimation smoothing operator for efficient incorporation of index data into prediction models
- Operator operates separately from original predictor, applicable to various machine learning tasks without retraining models
- Theoretical analysis establishes conditions under which post-estimation smoothing enhances prediction accuracy
- Experiments on large-scale spatial and temporal datasets demonstrate effectiveness and speed in practice
- Approach significantly improves prediction accuracy by incorporating natural structure of index variables
- Research offers new perspective on integrating structural index data into machine learning algorithms

## Summary

In their paper titled "Post-Estimation Smoothing: A Simple Baseline for Learning with Side Information," authors Esther Rolf, Michael I. Jordan, and Benjamin Recht explore the significance of natural structural indices in observational data, such as time stamps and geographic locations. These indices are often overlooked in prediction tasks, but the authors propose a novel approach to leverage them while ensuring resilience against potentially uninformative or misleading indices. Their key contribution is the introduction of a post-estimation smoothing operator that efficiently incorporates structural index data into prediction models. This operator operates separately from the original predictor, making it applicable to a wide range of machine learning tasks without the need for retraining models. Through theoretical analysis, the authors establish simple conditions under which post-estimation smoothing can enhance prediction accuracy compared to the original predictor. They also conduct experiments on large-scale spatial and temporal datasets to demonstrate its effectiveness and speed in practice. The results showcase how this approach significantly improves prediction accuracy by incorporating the natural structure of index variables in machine learning tasks. Overall, their research sheds light on a new perspective for considering and integrating structural index data into machine learning algorithms. This offers valuable insights for future developments in this field.