# Toward Unsupervised Outlier Model Selection

**Authors:** Yue Zhao, Sean Zhang, Leman Akoglu  
**arXiv:** [2211.01834](https://arxiv.org/abs/2211.01834)  
**Date:** 2022-11-03  
**Summary page:** https://www.summarizepaper.com/en/paper/2211.01834v1/

## Key points

- The paper addresses the problem of unsupervised outlier model selection (UOMS)
- The authors propose a new approach called ELECT for selecting an effective candidate model for outlier detection on a new dataset without labels
- ELECT leverages prior knowledge from similar historical datasets using meta-learning
- ELECT uses a performance-based dataset similarity measure to find similar historical datasets
- ELECT can adaptively search and provide output on-demand, suitable for varying time budgets
- Extensive experiments show that ELECT outperforms various UOMS baselines significantly
- The paper provides implementation details and code availability on GitHub
- Overall, ELECT offers a promising solution that surpasses existing baselines in terms of accuracy and efficiency

## Summary

The paper titled "Toward Unsupervised Outlier Model Selection" addresses the problem of unsupervised outlier model selection (UOMS), which is an understudied area despite the abundance of outlier detection algorithms in the literature. The authors propose a new approach called ELECT to select an effective candidate model, including its hyperparameters, for outlier detection on a new dataset without any labels. ELECT is based on meta-learning and leverages prior knowledge from historical datasets that are similar to the new dataset. By transferring information such as model performance, ELECT aims to facilitate UOMS. One unique aspect of ELECT is its use of a performance-based dataset similarity measure, which is more direct and goal-driven compared to previous measures used in this context. To find similar historical datasets, ELECT adaptively searches and can provide output on-demand, making it suitable for varying time budgets. The authors conducted extensive experiments to evaluate ELECT's performance against various UOMS baselines. These baselines include not performing any model selection (always using a popular model like iForest) and more recent strategies based on meta-features. The experimental results demonstrate that ELECT outperforms a wide range of basic UOMS baselines significantly. This highlights the effectiveness of ELECT in selecting outlier detection models for unlabeled datasets. The paper also provides additional details about the implementation and availability of code on GitHub. Overall, this paper presents an innovative approach to address the critical problem of unsupervised outlier model selection. By leveraging meta-learning and performance-based dataset similarity measures, ELECT offers a promising solution that surpasses existing baselines in terms of accuracy and efficiency.