1
$\begingroup$

My dataset contains 13 attributes consisting of 10 Numerical and 3 Categorical attributes and Target. It has 180 observations

NumFeature1,NumFeature2....NumFeature10,CatFeature1,CatFeature2,CatFeature3, Target

All Categorical are non-ordinal and each have the following categories:

CatFeature1: 0/1

CatFeature2: 0/1/2

CatFeature3: 0/1/2/3

It is a binary classification problem where we have to predict the probability for each class of the target class.

I have 3 Questions for above dataset:

Q1- For the categorical Feature, Should I use LabelEncoder() or OneHotEncoder() or df.get_dummies() or should I just combine custom label encoder with one-hot encoder?

Q2- Should scaling be done for Numerical features only or it should be done for all the features including categorical after encoding

Q3- What should be the best model to get the probability of the binary classification. So far, I have tried kNN, LogisticRegression, and RandomForestClassifier with predict_proba, but log_loss score was 0.301 the best.

$\endgroup$
  • $\begingroup$ Also, it would be helpful if someone could help me with the order of steps to be followed in data preparation for such type of datasets. $\endgroup$ – Ammar Akhtar Sep 16 '19 at 18:00
  • $\begingroup$ Unless asked for clarifications, please do not use the comments space to complement your question; edit & update your post instead. $\endgroup$ – desertnaut Sep 17 '19 at 23:26
1
$\begingroup$

Ans 1: get_dummies() or (label encoder + one-hot encoder) would do the trick.

Ans 2: Scaling categorical dummy data does not make sense. It also loses out on interpretability.

Ans 3: Logistic regression might tend to overfit since you only have 180 observations. KNN might perform well with a small number of observations but it doesn't handle categorical variables well. Random Forest or Extreme Random Forest might be your best bet. However, there's no "one fits all" concept in ML. You'll have to try a variety of algorithms and see which fits your dataset best. Though there are Linear SVM and other algorithms, I'd suggest that you try simpler ones as otherwise it might be an overkill.

Good luck!

|improve this answer|||||
$\endgroup$

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.