\(\DeclarePairedDelimiterX{\Set}[2]{\{}{\}}{#1 \nonscript\;\delimsize\vert\nonscript\; #2}\) \( \DeclarePairedDelimiter{\set}{\{}{\}}\) \( \DeclarePairedDelimiter{\parens}{\left(}{\right)}\) \(\DeclarePairedDelimiterX{\innerproduct}[1]{\langle}{\rangle}{#1}\) \(\newcommand{\ip}[1]{\innerproduct{#1}}\) \(\newcommand{\bmat}[1]{\left[\hspace{2.0pt}\begin{matrix}#1\end{matrix}\hspace{2.0pt}\right]}\) \(\newcommand{\barray}[1]{\left[\hspace{2.0pt}\begin{matrix}#1\end{matrix}\hspace{2.0pt}\right]}\) \(\newcommand{\mat}[1]{\begin{matrix}#1\end{matrix}}\) \(\newcommand{\pmat}[1]{\begin{pmatrix}#1\end{pmatrix}}\) \(\newcommand{\mathword}[1]{\mathop{\textup{#1}}}\)
Classifier Errors
Combined Orders
Multiobjective Optimization Problems
Needed by:
Sheet PDF
Graph PDF

Optimal Classifiers




Given two classifiers $G_1$ and $G_2$ and a dataset, we can associate to each its false positve and false negative rate on the dataset. For a finite dataset, these are two rational numbers. It is natural to prefer $G_1$ to $G_2$ if the former has a smaller false positive rate. Conversely, it is natural to prefer $G_2$ to $G_1$ if the former has a smaller false negative rate. Unfortunately, one may need to trade-off these two desiderata (see Combined Orders), since there is no total order. In other words, choosing betwen $G_1$ and $G_2$ is a multiobjective optimization problem.


Let $\mathcal{G} $ be a set of classifiers and let $f: \mathcal{G} \to \R ^2$ be defined so that $f_1(G)$ is the false negative rate of $G$ on some dataset and $f_2(G)$ is the false positive rate of $G$ on the same dataset. The $\kappa $-scalarized error metric (or Neyman-Pearson metric associated with $G \in \mathcal{G} $ is $\kappa f_1(G) + f_2(G)$. In the case that $\kappa > 1$, false negatives are given higher cost than false positives, and vice versa whtn $\kappa < 1$. For $\kappa = 1$, the scalarized error metric is the same as the overall error rate.

  1. Future editions will include. ↩︎
Copyright © 2023 The Bourbaki Authors — All rights reserved — Version 13a6779cc About Show the old page view