I'm trying to work on this. I have a little question, however.
I don't really understand the definition of TP, TN, FP, FN when the classification type is not binary.
I mean, what is "positive" when it is multiclass, or multilabel?

By the way, sensitive is by definition TP / (TP + FN). I'm wondering what if TP + FN = 0?
That is, when y_true = [0, 0, 0, 0]?

I'm not really familiar with the terms TP, FN etc. So please correct me if I make any mistake.

Thanks.

Perhaps it should just support the binary case at first.

On 31 July 2014 03:38, Bert Chang notifications@github.com wrote:

I'm trying to work on this. I have a little question, however.
I don't really understand the definition of TP, TN, FP, FN when the
classification type is not binary.
I mean, what is "positive" when it is multiclass, or multilabel?

Thanks.

—
Reply to this email directly or view it on GitHub
#3506 (comment)
.

TP, TN, FP and FN are only well-defined when one class is considered negative (which is quite common). Usually we pick the class that is smallest according to Python's standard ordering unless overridden by the user.

@lazywei , did you get a chance to start on this? If not, I may take a stab at it.

@adam-m-mcelhinney
I have started a PR at #3511 , and I thought I've finished...
Anyway, I think you could check out my branch to do more base on it if you want. In that case, let me know when you have done, I'll close my PR.
Thanks

Got it. Thanks.

@ppuggioni you have worked on this during the sprint: can you please open a PR with a [WIP] marker in the title with the current state of your work?

@ogrisel yes, but yesterday I realised that @lazywei had possibly already done the job. I will push mine and open a PR with [WIP], but I was thinking that it might be worth for me to compare mine to his to avoid duplications and double review process?

Indeed, please submit yours and cross reference the 2 PR to compare the results.

Hi all,

May I take over this issue? It seems that the referenced two PRs have stalled.

I have done coding, documentation, and testing. I found the balanced accuracy score is equal to the average of positive label recall and negative label recall. I did not take the balance weight into account. Since in the wikipedia, it is fixed as 0.5. So my code is quite simple. May I open a [WIP]PR? @ogrisel @jnothman

I am new to scikit-learn development, but I have used the package for a long time. I am a PhD student studying in machine learning. I hope I can join in GSOC 2015.

Thanks,

I haven't touched it. I believe someone else had finished it however.

On Wed, Feb 25, 2015 at 4:20 PM, Wei Xue notifications@github.com wrote:

Hi all,

May I take over this issue? It seems that the referenced two PRs have
stalled.

I have done coding, documentation, and testing. I found the balanced
accuracy score is equal to the average of positive label recall and
negative label recall. I did not take the balance weight into account.
Since in the wikipedia, it is fixed as 0.5. So my code is quite simple. May
I open a [WIP]PR?

I am new to scikit-learn development, but I have used the package for a
long time. I am a PhD student studying in machine learning. I hope I can
join in GSOC 2015.

Thanks,

—
Reply to this email directly or view it on GitHub
#3506 (comment)
.

hi, @adam-m-mcelhinney,

Searching keywords balanced accuracy shows only two relevant PR #3929 and #3511. They have no progress since last year. If some else had finished it, he/she would have cross reference this issue #3506.

Looks like its all yours then. Let me know if you want to collaborate on
anything.

On Wed, Feb 25, 2015 at 5:09 PM, Wei Xue notifications@github.com wrote:

hi, @adam-m-mcelhinney https://github.com/adam-m-mcelhinney,

Searching keywords balanced accuracy shows only two relevant PR #3929
#3929 and #3511
#3511. They have no
progress since last year. If some else had finished it, he/she would have
cross reference this issue #3506
#3506.

—
Reply to this email directly or view it on GitHub
#3506 (comment)
.

Closed by #8066.