effective sample size 相关 AI 资讯

10:02

官方账号arXiv cs.LG@Adam Noonan

This paper discusses the challenges in setting thresholds in machine-learning systems, particularly when calibration examples are not independent. It introduces a new approach to calculate the effective sample size for thresholds under clustering, showing that the current correction in the conformal literature is incorrect. The study also highlights the importance of considering the threshold level for each dataset, as the effective sample size varies with the threshold setting. The research is based on a calibration set of 25,028 examples and measures the reliability of about 1,300 thresholds.

论文 thresholds machine learning calibration sets

推荐理由：This paper offers a fresh perspective on threshold setting in machine learning, providing a new method to calculate effective sample sizes under clustering. It's a must-read for those interested in improving the reliability of machine learning systems.

原文