0
$\begingroup$

I am working with a binary classifier and I want to express the "balance" or "skewness" of the training data using a metric.

I want to reflect this ratio in a report, like this:

Accuracy: 80%
Recall: 78%
Precision: 62%
*The Ratio of Positive to Negative Samples*: 62%

I feel like there's probably a standard name for the metric expressing "The Ratio of Positive to Negative Samples".

My main question is: What is the name of this metric?

(This question assumes that there is a standard name for this metric.)


Extra Information:

Here are some example values for the metric and their interpretations:

1.0 = The sample (of training data) is balanced.
0.5 = There are twice as many negative samples as positive samples.
2.0 = There are twice as many positive samples as negative samples.

My secondary question is: Is there another metric that can be used to form similarly meaningful interpretations?

$\endgroup$
2
$\begingroup$

You can use class ratio / sample class ratio. Which will make it more intuitive for any reader while going through the details.

As its not used for model performance analysis hence I think we don’t have a metric name for this.

| improve this answer | |
$\endgroup$
  • $\begingroup$ Class Ratio may be the best answer so far. Thank you. :) $\endgroup$ – Eric McLachlan Dec 4 '19 at 9:13
  • 1
    $\begingroup$ Thank you..Sometimes we forget the most basic things amongst the heavy machine learning jargon. $\endgroup$ – cap Dec 4 '19 at 15:35
1
$\begingroup$

This is Imbalance Class Dataset concept. Mostly the ratio you mentioned is used as:

# of positive class (minority) / # of negative class (majority)

For example: "The dataset contains 100 fraud activities among 10000 transactions with 0.01 class imbalance."

There are no strict rule/metric about that you can also use other version with 100 class imbalance but its not preferred at all.

Class Weights should be the term you looking for.

| improve this answer | |
$\endgroup$
  • $\begingroup$ I was hoping there would be a single-word name for the metric like "recall", "sensitivity", etc. Do you know if this metric has a name? Anyway, thank you for your answer. $\endgroup$ – Eric McLachlan Nov 25 '19 at 12:38
  • 1
    $\begingroup$ as far as i know class weights (as a pre model metric) the term you looking for. Recall, sensitivity are performance metrics - post model metric. $\endgroup$ – Ilker Kurtulus Nov 25 '19 at 12:47
  • 1
    $\begingroup$ I don't think this is correct answer, class weight is I believe used to weight losses, so say the loss incurred for misidentifying the result is weighted according to the weighting value you set. It just happened that most of the time the weighing value uses the ratio between minority and majority class. $\endgroup$ – Yohanes Alfredo Nov 25 '19 at 13:10
  • 1
    $\begingroup$ Weighting "coefficients" to minimize loss and "class" weights (simply ratio) are different things. $\endgroup$ – Ilker Kurtulus Nov 25 '19 at 13:15
  • 2
    $\begingroup$ Have you tried Entropy? (en.wikipedia.org/wiki/Entropy_(information_theory)) It expresses what you are looking for, it is not designed for how you want to use it. $\endgroup$ – S van Balen Nov 25 '19 at 13:31
1
$\begingroup$

It's similar to the concept of odds. The baseline count of positive and negative samples gives you an odds (all else being equal) that a random sample from the population will be positive or negative.

| improve this answer | |
$\endgroup$
  • $\begingroup$ Thanks, @Tom. I've rewritten the question to better indicate my use case. $\endgroup$ – Eric McLachlan Dec 4 '19 at 7:20
0
$\begingroup$

There is no standard, generally accepted label for this metric.

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.