2
$\begingroup$

After applying a machine learning algorithm on some data, I have a table with one column being the prediction value (can be either binary or continuous). The other columns are the independent features (mostly numeric values). I'm trying to look for a way of visualization which answers the question: Which of the independent features have the most correlation with the prediction value? I tried to find something but couldn't find anything; it might be though that I'm not looking for the right terms. Please excuse if my question lacks clarity as I'm still new in the realm of machine learning.

$\endgroup$
  • 3
    $\begingroup$ Side note: be aware that correlations ( / the lack thereof) can be largely driven by the extent ( / the lack thereof) of the range of the features. This fact may mean that stronger ( / weaker) correlations don't imply what people will intuitively assume. $\endgroup$ – gung Jan 2 '13 at 19:54
  • 2
    $\begingroup$ It may well depend on the kind of classifier you used. E.g., if it relies on combination of features or complex measure of variable importance or some non-linear mapping, such 'univariate' visualization approach (i.e., feature vs. outcome) won't probably bring any insight. $\endgroup$ – chl Jan 2 '13 at 20:40
  • 1
    $\begingroup$ Looking at the data is great. But if you've already made a model, then it might be more fruitful to figure out how to extract feature importance from your model and visualize that. Also, you might mention how big your data set(s) get (samples, features) since not every technique scales to large sets. $\endgroup$ – MattBagg Jan 2 '13 at 20:41
  • $\begingroup$ +1 to each of the comments by gung, chl and MattBagg. Bearing their points in mind, though, have you started with a basic series of scatter plots with a feature on the horizontal axis and the prediction value on the vertical? This would only be a starting point. $\endgroup$ – Peter Ellis Jan 3 '13 at 22:42

Your Answer

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

Browse other questions tagged or ask your own question.