1
$\begingroup$

I have tried Box-Cox , exp, log etc but some features are not converting into normal distributions Please suggest me some alternative option ..

As you can see in second and third graph it is not Normal ! but 1st and 4th is normally distributed I use this equation to convert into Normal form 1/(max(df[fetaure_name] + 1)- df[feature_name]) I have also tried other methods as I mentioned above.
Adding more details : details of second graph :

count    1867.000000
mean      177.746754
std        10.191609
min       170.776654
25%       174.561455
50%       175.982265
75%       178.388291
max       357.000475
mode --> every value is unique 

enter image description here

**original distribution of second feature ** enter image description here

$\endgroup$
  • 1
    $\begingroup$ Your histograms don't show enough detail to allow confident advice. Key questions are: do any of those variables include exact zeros? Do any of those variables show spikes in their distribution? Over and above that, what is it that you want to do that you think requires marginal normal distributions? I see that you have about 4000 observations, but you should be able to show minimum, median, quartiles, maximum, mean, SD for those 4 variables. $\endgroup$ – Nick Cox Apr 13 at 11:58
  • $\begingroup$ I really wouldn't call your first and fourth distributions normal. It seems that there is a bound at 0 and perhaps a spike of 0 of very low values. Also, are values bounded by 0 and 1? If so, the normal isn't even a reference distribution. $\endgroup$ – Nick Cox Apr 13 at 12:05
  • $\begingroup$ @NickCox i have added more details in original question , yes 1st and 4th graph is not extactly normally distributed but i am looking for something which looks or more near to normal distribution . i have added detail for second graph .please check once , thanks $\endgroup$ – user3219871 Apr 14 at 11:30
  • $\begingroup$ Thanks for adding detail, but what you added enhances the mystery here. If the range of one variable is from about 171 to about 357 the histogram limits that appear to be 0 and 1 are misleading. We can't advise on sensible transformations without knowing the real range of the data, including whether variables are ever zero (or negative). The best method of assessing approximation to normality in my view is normal quantile plots for the original data. Most of the questions in my first comment are not yet answered, so this is too unclear without much more detail. $\endgroup$ – Nick Cox Apr 14 at 11:40
  • $\begingroup$ Your code in some unstated language 1/(max(df[feature_name] + 1)- df[feature_name]) seems to mean $[1/ [\text{max}(x + 1) - x]$ for a variable $x$, which needs more explanation. $\endgroup$ – Nick Cox Apr 14 at 11:43

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.