For which set of values of predictors (numerical) will my response variable (factor) be guaranteed to equal a certain level (using Random Forest)

I am going to explain it on base of the iris dataset:

Goal --> I need to get as many species of setosa as possible

The way --> i must find the best (representative) combination of: "Sepal.Length","Sepal.Width","Petal.Length","Petal.Width" values, which identify only setosa and not any other species.

I would appreciate any help as i cannot find any hint or way how to handle this problem.

--> mean of the parameters set in the group of setosa is not sufficient! (as in my original dataset the difference between groups are not that clear!)

To get a bit background of my real problem is that i have different temperature sensors at different points in the production line, and i must identify which combination of those temperatures leads to good quality product, and which combination leads to bad quality product, i think it is very simple to understand. We must find best combination of temperatures and try to achieve it at the production.


I have found a similar question on cross validated (but my target value is a factor, not numeric as in this case, but the main princip is the same:

How would I be able to find for which values of inputs do i get the setosa target variable

), however it has not been answered.

  • 2
    $\begingroup$ What does it mean to "get" species of setosa? If you label all of your data "setosa," you're guaranteed to "get" all of the setosa in it (and a lot of non-setosa also). So it seems more sensible to "get" setosa and only setosa, which sounds like a classification problem, which is an enormous field. I recommend you start by reading Elements of Statistical Learning. $\endgroup$
    – Sycorax
    Jul 27 '17 at 14:41
  • $\begingroup$ Exactly i wanna "get" setosa and "only" setosa. I have tried already classification using RF, but i have no idea how i can get the best representative combination of values and not only predicition. I got really stuck at this point $\endgroup$
    – Mal_a
    Aug 2 '17 at 6:28
  • $\begingroup$ I think you are trying to get the ranges of values in the variables for which a typical outcome would be Species=setosa. Correct? In this case a descriptive statistics with confidence intervals would maybe suffice? $\endgroup$
    – Patrik
    Aug 2 '17 at 9:29
  • $\begingroup$ The classification model aside, unless you only have one predictor, there might be/are going to be multiple predictor value combinations which result in a certain classification. For example, a high value on predictor 1 on its own might result in (high probability of) class 1, while moderate to high values on predictors 2 and 4 might also result in class 1. My question/suggestion: do you want to find the single least 'errorprone' classification rule, or would you also want to combine such predictorpatterns into a classification rule? $\endgroup$
    – IWS
    Aug 2 '17 at 10:57
  • $\begingroup$ Well i do have more then one predictor (my original dataset has at least 3 predictors), therefore i need combination of all of them together $\endgroup$
    – Mal_a
    Aug 2 '17 at 11:00

It sounds like you're looking for the term classification. That entire field deals with models trying to predict a class (setosa or not) from some features.

  • $\begingroup$ Exactly i am looking for classification (i already did it with RF), however what i get from it is only predicition of the Species and i need to find the representative values set (the best combination) of parameters which lead to predicition of one of the Species, in my case it is setosa.As for example values of: Sepal.Length 5.1, Sepal.Width 3.0, Petal.Length 1.4, Petal.Width 0.2 are most representative of Species Setosa. $\endgroup$
    – Mal_a
    Aug 2 '17 at 6:26
  • $\begingroup$ Then you should search for answers (or post a new question) specifically about the method you used, because this process is very different for different approaches. Also, you will need to define your loss function - i.e. what do you mean by "most representative"? Identify only setosa? Identify as many setosa as possible within some predictor bin? Make least classification errors?.. $\endgroup$
    – juod
    Aug 2 '17 at 8:38

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