1
$\begingroup$

I have a dataset with more than 100,000 observations (rows) and 24 variables in which 23 are continuous and one is categorical variable.

The categorical variable has 13 categories (1, 2, 3, ..., 13) and one more category (0) which is outlier.

I have to build a model which will predict outliers (category with 0) with the highest accuracy.

Should I apply k-mode clustering? Or which algorithm will be most suitable?

I was thinking of combining all other categories except 0. Because 0 are outliers and ultimaate objective is to find those outliers. Is this good approach?

Improve this question
$\endgroup$
2
  • $\begingroup$ If the goal is to detect outliers it makes sense to merge the remaining classes. I wonder however, how are those outliers defined? If they are defined based on the given data, then I don't think it makes any sense, as the rule used to classify something as an outlier could directly be used. $\endgroup$
    – George
    Commented Jul 24, 2016 at 13:08
  • $\begingroup$ @George i dont know about the data.. i mean whether its a sensor data or its a machine data or data collected randomly. i am not sure. Its just have high dimensions and my goal is to find outleirs with high accuracy. $\endgroup$
    – kush
    Commented Jul 24, 2016 at 13:27

1 Answer 1

Reset to default
3
$\begingroup$

When in doubt, and when you don't have good reasons to choose otherwise, grow a Random Forest, because Fernández-Delgado et al. (2014). I often find it hard to improve on RFs. However, do think about the trade-offs between false positives and false negatives, or sensitivity and specificity.

References:

Improve this answer
$\endgroup$
2
  • $\begingroup$ I was thinking of combining all other categories except 0. Because 0 are outliers and ultimaate objective is to find those outliers. Is this good approach? $\endgroup$
    – kush
    Commented Jul 24, 2016 at 12:22
  • $\begingroup$ That may indeed make sense, if you are not interested in the other categories. Best to try both methods and see how they perform on a holdout sample. $\endgroup$
    – Stephan Kolassa
    Commented Jul 24, 2016 at 14:31

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.