Lec9 - School of Computer Science and Technology University...

Info iconThis preview shows pages 1–5. Sign up to view the full content.

View Full Document Right Arrow Icon
Nature Inspired Computation and Applications Laboratory School of Computer Science and Technology University of Science and Technology of China Pattern Recognition Lecture 9 Output Calibration and Evaluation of PR Models
Background image of page 1

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
Nature Inspired Computation and Applications Laboratory 主要内容 校准分类器的输出 分类器的评测指标 估计分类器的推广性能( generalization performance
Background image of page 2
Nature Inspired Computation and Applications Laboratory 校准分类器的输出 对 于一 个 待分 类 的样 本 ,任 何 分类 模 型 ( 如 Parzen 窗方法、 k- 近邻、线性判别函数、神经网络、 判定树等)判定该样本类别的本质可以理解为对 该样本属于不同类别的“可能性”进行打分,然 后将其判定到得分最高的那一类。 在实际使用中,我们往往希望这些“可能性”直 接表现为样本属于某一类别的后验概率,这样便 于我们利用一些概率方法寻求最优的判定函数。 但是,许多模型(如神经网络、 SVM )的输出本 身并不满足这一要求,而某些模型的输出虽然满 足概率的一般条件,却可能很不准确。这就要求 我们对分类器的输出进一步校准。
Background image of page 3

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
Nature Inspired Computation and Applications Laboratory 校准分类器的输出 对于一个 2 类分类问题,假定我们已经通过某种分 类模型对 n 个样本进行了打分( score )。例如, 对于 SVM ,我们获得一个 n 维向量。 考虑一个样本属于正类的后验概率,真实的概率 是对于所有正类样本,概率为 1 ,否则为 0 若我们希望将上述 score 校准为后验概率的形式, 则本质上是需要在 score 和真实的后验概率之间寻 找一个合适的映射,使得 score 在映射后尽可能地
Background image of page 4
Image of page 5
This is the end of the preview. Sign up to access the rest of the document.

This note was uploaded on 06/16/2011 for the course CS 1256 taught by Professor Tangke during the Spring '10 term at USTC.

Page1 / 14

Lec9 - School of Computer Science and Technology University...

This preview shows document pages 1 - 5. Sign up to view the full document.

View Full Document Right Arrow Icon
Ask a homework question - tutors are online