Although the tfidf vectors don't really follow Multinomial Distribution, yet MultinomialNB works fairly well, why is it so? Also would weighted tfidf w2v work the same way or should I use GaussianNB for that? plus since w2v vectors can be negative too, so it makes sense to use minmax scaler on it. but this would then affect be affected in presence of outliers. Also i had read somewhere that doing processing on text vectors can be actually bad because they may not represent the actual text then.