Would suggest:

1. Have a look at the number of samples from each class ;

2. Identify outliers. i.e., any class that has only a few samples, for example, less than 1%;

3. Depending on the context of the data, remove the outliers, or merge them into a class, say, class "other";

4. Randomly select samples from each class, so both the training and test data sets will have samples from the same classes.

5. Deal with imbalanced classification problems.

In addition, you may not need to do one-hot encoding - depending on which model you use, for example, decision trees.

You can also consider dimensionality reduction - do classification in a low dimensional space.

Hope this makes sense and helps.

A classifier cannot classify the samples of a class if some samples of the class aren't present in the training set. It's like one won't be able to recognize a certain species of birds if he hasn't already seen some birds of that species before, and marked the distinctive features that he will look for in a new bird to categorize it.

A good approach to solve your problem can be building a customized training set. Instead of splitting the whole set of samples randomly to train and test sets, split the samples of each class (randomly) separately (preferably 70%-30%) and then construct the train and test sets by joining the subsets of each class. This process will ensure that there is no class present in the test set which is unknown to the classifier. Hope this helps.

Hello Kalyan

1. Create a separate file (.txt) for labels and and mark them accordingly

2. Follow the code as in attach file may help you to figure out your question.

Regards