# When to choose Regression techniques in supervised Learning?

I have two Questions regarding when to choose Classification and Regression methods.

1. In my experiment, 'target' can be coded either by one of the two features, one feature is numerical discrete and other numerical continuous. In my set of experiments the numerical continuous feature is rounded off to nearest round figure but its time so its continuous. It can be coded by both individually, its upto me which feature to use to code my target (I will code by one of the features). Then will it be regression case or classification in supervised machine learning?

2. But what if i chose other feature to code my target, the discrete feature, then can I apply regression to it ?

As from what I know so far, Regression methods are applied when response being predicted is ordered and continuous. (I dont know more detail about it)

• is the discrete feature ordered? Nov 27 '15 at 11:47
• @highBandWidth the discrete feature is frequency of occurrence of an event. (will that be ordered or not ? )
– Umar
Nov 27 '15 at 11:50
• Yes,since you know two occurrences are more then one occurrences and one and two occurrences are closer to each other than one and 20 occurrences. Nov 27 '15 at 11:52

Since both your targets are ordered, you should use regression. For the discrete target, you can use the ordered probit or ordered logit model.

• there is confusion, I have two features and can use either of two to code my target. So I will use one to code only. The target will be either 0 or 1. So from your response I understood that as I have ordered discrete even if I select that to code my target I can use regression.. Is that right?
– Umar
Nov 27 '15 at 12:42
• @Umar, yes, you should use regression whichever target you use. Nov 27 '15 at 14:37

Classification methods are generally for problems such as distinguishing between red, blue and green, benign vs. malign or male vs female. That is, things that are distinct categories. Regression is (generally) for things that can be measured using numbers, such as time or magnitude.

So yours is a regression problem.

Since your target is a duration/time, you should look at duration regression models such as the cox proportional hazards model.

• my Target is not time, one of the features is time, I want to code my target to boolen value depending on either of the two features.
– Umar
Nov 27 '15 at 12:36
• If your target variable is boolean, then you are looking for a classification model such as probit og logit regression, decision trees, random forests, support vector machines et cetera. You should not use ordinary linear regression to predict a boolean variable. Nov 27 '15 at 14:25