I have been working on an email data set, and trying to predict the owner team for it. But my prediction accuracy is just 58%. I have implemented data cleansing, null value removals, duplicate removal, excluding stopwords, and then calculating tf-idf value to get my final document term matrix.

Tried various classifiers (but with default settings) but max I achieved is 58% accuracy. Might give a try to hyperparameter tuning.

Could you please give me some pointers, based on your previous experience what possibly could be going wrong? I'm simultaneously doing my own research and going through various blogs and articles.

One thought - Could it be a case of data skewness?

My output classes are distributed like this - enter image description here

Your thoughts please?


Just telling the accuracy does not mean anything in classification problems. The first thing you must do is to calculate your baseline, that is, what is the percentage of your majority class? For the bar plot above it is difficult to measure, can you tell us in percentage, instead of counts? In this way we can assess better your results.

Also, have you plot a confusion matrix? In this way you can see where your model are getting more wrong and try to infer why this is happening.

And yes, since you have too many classes to predict and most of them are with low representativeness this will be difficult to overcome. Maybe you can try things such as Oversampling, Undersampling techniques considering one-vs-all approach. This is just an idea, I haven't yet encountering a problem with so many classes to predict.

  • $\begingroup$ Thanks Victor. I read few articles which introduced me to Oversampling and Undersampling, SMOTE, etc., but that's entirely new for me. Will do more research and get back with specifics. I'll amend my question by adding the Confusion matrix in sometime. $\endgroup$ – ranit.b Mar 6 '19 at 19:01

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.