0
$\begingroup$

Data Structure:

Method Category   Variance for X
1         A             20
1         B             14
1         C             16
2         A             14
2         B             19

Where X was not used for classification, but is evaluation criteria. The objective is select method which produces classification with minimum possible variance for X for most of class / overall least variance.

My question: is there some standard (or obscure) method of visualizing variance for large number of categories?

$\endgroup$
  • $\begingroup$ How "large" is the number of categories? $\endgroup$ – Romain Reboulleau Sep 29 at 6:46
  • $\begingroup$ 4 Methods. 12 Categories. $\endgroup$ – Martan Sep 29 at 7:05
0
$\begingroup$

I know this may not exactly correspond to what you need, but my first idea would not be to visualize the variances. Instead, I would define a metric for each method.

For instance, considering all results obtained with a given, you have a vector of outputs (one Variance for X for each Category), then you can simply compute a p-norm of this vector, and compare it to the norm obtained for other methods:

$$\left\lVert x\right\rVert_p = \left( \sum_{i=1}^n x_i^p \right)^{1/p} $$

$p=2$ gives you the euclidian norm, you can increase or decrease $p$ depending on if you want to raise attention on maximal values or the average.

In terms of visualization, you could simply plot (in a bar plot for instance) a few norms that you selected (orders 1, 2, and infinite for instance).


For a pure visualization, you could plot the data in violin plots. For instance in python:

import seaborn as sns
sns.violinplot(x="Method", y="Variance for X", data=your_data_as_df)

Each violin would give you an idea of how the data is distributed over all categories.

$\endgroup$

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.