85.txt

This complete matlab for neural network

发信人: freehorse007 (freehorse007), 信区: DataMining
标  题: Re: 请教关于宏平均和微平均
发信站: 南京大学小百合站 (Mon Apr  7 19:19:37 2003)


This issue is addressed in the following paper:

@InProceedings{Yang:2001:STS,
  author       = "Yang, Yiming",
  title        = "A Study on Thresholding Strategies for
                  Text Categorization",
  booktitle    = "Proceedings of the 24th International Conference
                  on Research and Development in Information Retrieval",
  year         = 2001,
  month        = "September",
  pages        = "137--145"
}

Here is an excerpt from Section 4.1, page 140:

  It may be worth mentioning a potential confusion in
  computing macro-averaged F1 . That is, there are two
  possible ways: the "correct" way, according to our denition,
  is to compute the F1 values for each category and then take the
  average over the per-category F1 scores; the "wrong" way
  is to first compute the macro-averaged recall and macro-averaged
  precision, and then compute the harmonic average of these averaged
  recall and precision. The two ways often yield different results:
  the "correctly" computed F1 value is often signicantly lower than
  the "wrongly" F1 value because the F1 value of a specific category
  is usually dominated by the smaller value between the recall and
  precision values for that category when these two values are
  radically different.

--

※ 来源:．南京大学小百合站 http://bbs.nju.edu.cn [FROM: 202.117.85.146]