4 Class Classification - Machine Learning Model

Question

I have a data set which contains nearly 150 features and 60k data. And my target feature is continuous variable represents hours. I divided this period into 4 categories of user engagement (4 ranges of hours). Implement GA with SVM, GA with logistic regression, Random forest, GA with KNN with suitable normalization of data wherever required. Used GA for best feature subset selection.

All the algorithms gave similar results of around 46% accuracy ( for nearly balanced test set).

Note: Training is also done on a balanced data set. I am wondering where am I gone wrong?

I believe I went wrong somewhere in input to target mapping. Could anyone please confirm that categorization of the continuous target variable (hrs) into 4 sets are reasonable?

Answer 1

Yes, binning a continuous variable so that in can take discrete values is reasonable, as long as you are OK with transforming this regression problem into a classification one.

Just note that when dealing with a balanced 4-class classification problem, if an algorithm hadn't learnt anything and was predicting randomly it would achieve a 25% accuracy. Your results indicate that the algorithms are at least learning something. It might just be a very difficult task to solve, in that the features might have little to no correlation with the output variable.

I'd suggest trying out more algorithms with and without GA.