1
$\begingroup$

Having issues with the "find informative projections" feature in Orange 3. In order to be able to use this feature in the scatterplots, you have to select a target variable in the "select columns". So I choose the value that I'm most interested in finding correlations to. However, whatever I choose as a target variable doesn't show up in the informative projections list as being correlated to other features in my data set. Why not?

I have a lot of features in my data and was under the impression that the "find informative projections" option finds pairs of features that seem to have a linear correlation and sorts them by which pairs are most correlated.

If that's the case, why do you need a target variable? How do I choose an appropriate target variable?

Am I misunderstanding the purpose of the "find informative projections" feature?

I'm attaching a pic of my work flow. Orange data workflow

$\endgroup$
0
$\begingroup$

As you said, "find informative projections" gives you the best pair of features ("score plots") to explain the target variable. The two best features will be on the x- and y-axis, while your target variable will be the color (hue for numeric, categorical otherwise).

Here is an example for the Iris dataset. Where "iris" is the target --> color, and petal length and petal width are the most informative features, followed by petal width and sepal width, and so on.

enter image description here

| improve this answer | |
$\endgroup$
  • $\begingroup$ Thank you. I've seen that example before. What I don't understand is why if iris is the target variable it does not show up in that list as being correlated to petal length or sepal width, etc. I was hoping it would give me a list of the top pairs of features that are highly correlated, or have the highest r-squared value or something. This seems like it's going one step deeper. What kind of analysis is it doing to come up with these projections? $\endgroup$ – buske Sep 5 '17 at 16:46
  • $\begingroup$ for a given target $y$ the it gives you the best two features $(x_j,x_k)$ that predict $y$. If you don't want tuples of features, you can use the rank widget for regression or the preprocessing widget with select relevant features for classification $\endgroup$ – oW_ Sep 5 '17 at 17:41
1
$\begingroup$

Slightly related with the question (issues with Find Informative Projections), I found this post trying to solve this issue: the tool Find Informative Projections was disabled and Orange did not provide any hint about why.

In the end the issue was on my side: I forgot to set any variable as Target when loading the data so it was useless to select it in the Scatter Plot points-colors section.

Double click on the File to load the data where you can choose if a variable is categorical or numerical, then double click on the variable I wanted to use as Target (that was marked as Feature) in order to select it as Target and the Find Informative Projections was enabled again.

PS: I have also noted that when making predictions, the file loaded should have all variables marked as Features. If you mark as before one of the variables as target, Orange shows a red circle with an exclamation mark on the Predictions widget, and a message: "Data does not have the same target as predictors" (even though it is). I guess that the target in the prediction file is automatically set by checking which info is missing in the file.

| 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.