2
$\begingroup$

I have data that defines the characteristics of an elevator. This data contains hundred of fields (height, weight, speed, number of persons inside the elevator, etc...). From this data, I want to classify the elevators according to their complexity. For example, I want to find which elevators share the same properties even if they are slightly different in some characteristics.

To tackle this problem, I was thinking of finding the most important features of the elevator first and later group them according to the results. But I don't know if I should use Random Forest to find the most important features so that I could group them with the help of K-means.

Should I use Random forest to find the most important features of an elevator? Or, should I use any other model like feature importance or correlation map?

How to know which statistical model should I choose?

Improve this question
$\endgroup$
1
  • $\begingroup$ have hundreds of features is pretty normal and people usually will not call it a huge dataset. $\endgroup$
    – Haitao Du
    Commented Feb 19, 2020 at 13:24

2 Answers 2

Reset to default
4
$\begingroup$

The question you may ask first is what defines "important feature".

Random forest is a supervised learning algorithm, you need to specify a label first then the algorithm will tell you which feature is more important respect the given label. In other words, specifying different label will have different results for variable importance.

Without using the label, algorithm such as PCA will define the variable that have large variance is important, which is another good starting point. This is intuitive because variable with large variance usually has more information and variable with zero variance means everyone is the same, and therefore this feature can be less useful.

Improve this answer
$\endgroup$
3
  • $\begingroup$ PCA selects directions (vectors) that account for maximum variation in data, which is a combination of the directions determined by the original variables. How does it define the variable that has large variance? $\endgroup$
    – naive
    Commented Feb 19, 2020 at 14:44
  • $\begingroup$ You can have a look at Andrew Ng's notes on PCA for very simple and concise explanation : google.com/… $\endgroup$
    – Camille Gontier
    Commented Feb 19, 2020 at 15:37
  • $\begingroup$ If PCA transforms several features of data into some principal components, how can we understand the output of PCA related to the input features? And if PCA outputs more than 2 PCs, how we interpret that? $\endgroup$
    – xeon123
    Commented Feb 19, 2020 at 16:51
-1
$\begingroup$

This looks like a good use case to use chi-squared feature selection. You can find more about its implementation in python and also some examples with documentation here. What this method does is it keeps only the features which have higher effect on your label

Improve this answer
$\endgroup$
4
  • $\begingroup$ chi-squared work with unsupervised data? please note that I have quantitative and categorical data. Do I need to remove categorical data to use chi-squared? $\endgroup$
    – xeon123
    Commented Feb 19, 2020 at 16:54
  • $\begingroup$ it does work those values, yes. Yet, it will not work with negative values.From your description of the problem, it seems to me that, you can also bin your continuous features. $\endgroup$
    – Emil Mirzayev
    Commented Feb 19, 2020 at 17:03
  • $\begingroup$ Can I put several features in the same bin? $\endgroup$
    – xeon123
    Commented Feb 19, 2020 at 17:08
  • $\begingroup$ You can bin them, but for every feature, you should have its separate bins. $\endgroup$
    – Emil Mirzayev
    Commented Feb 19, 2020 at 17:10

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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