I am looking for a general best practices regarding classification and correlations. I created a new predictor feature call it B, based on a certain threshold in a feature A. Now I started to do EDA and I am not sure which feature to include in my EDA, A or B. When I do correlations plots, nothing correlates with feature B, but some features do correlate with feature A. Which one should I take into account then, A or B correlations? Also, how can I make use of those correlations and scatterplots and pairplots anyway and are they important? If I am using random forest or NN, do I even need to bother with all of the pairplots and correlations to extract features from? I have around 150 features and not sure how to approach the problem of which features to use. I haven't found a source saying how to make a proper use of all of this in a real world scenarios. Any help is appreciated.

  • $\begingroup$ Use corr matrix to “filter” those highly correlated features and from those highly correlated features keep those least correlated to target. This is the best way to approach the problem. On top of all feature B is part of FE so that would be based on some domain knowledge and the reason you made B is not random … this is why no correlation is present (or you need to review that part or you should drop since no impact to the model would have) $\endgroup$
    – n1tk
    Aug 6 at 1:07

1 Answer 1


If the relation between predictors is nearly 0, it's always better to drop that feature, the caveat here depends on the domain knowledge you have.

Did you check the correlation between B and the target variable and also A and target variable? if it's negative drop it, If it's significantly high .i.e greater 0.7, use that as your feature.

Yes, pair plots and scatter plots are really important, but it would be tedious to plot features with 150 variables.

  • $\begingroup$ I don't have any domain knowledge, I have some random data with features that I don't know much about. How do I use scatterplots and pairplots to determine which features to use? Or what should I be using them for? $\endgroup$
    – user69194
    Mar 10, 2019 at 8:38
  • $\begingroup$ Do a correlation map or heat map to determine which features to use. negatively correlated features can be avoided, but with caution. Pair panels give a lot of information about your data points. You can also do a simple describe on your data and find about the skews. $\endgroup$
    – Sunil
    Mar 10, 2019 at 12:54

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.