I have 30,000 phrases categorized by sentiment.
I'm gonna use Naive Bayes.
Here's the proportion (sentiment -> number of phrases).
anger 98
boredom 157
empty 659
enthusiasm 522
fun 1088
happiness 2986
hate 1187
love 2068
neutral 6340
relief 1021
sadness 4828
surprise 1613
worry 7433
So, I have to split my dataset into train/test to execute my model, etc, right?
Should I keep the proportion of the categories when executing the stratification?
I mean, if I pick 30% for the test sample, should I keep 30% of each sentiment instead of 30% of the whole dataset?
I guess yes, but I would like to have a more experienced opinion.
And how would you do that? Anyone here know a better way of doing that instead of executing a python loop, testing which sentiment, calculate 30%, put in a dictionary etc?
Is there any Pandas trick to stratify by a category feature, keeping the proportion?
Should I keep the proportion of the categories when executing the stratification?
You seem a little confused regarding the terminology; the very definition of stratification (or stratified sampling) is exactly to maintain the proportions, otherwise it is simple random sampling.
if I pick 30% for the test sample, should I keep 30% of each sentiment instead of 30% of the whole dataset?
They are not contradictory, are they? If you keep 30% of each category, won't you end up with the 30% of your initial set?
Is there any Pandas trick to stratify by a category feature, keeping the proportion?
Don't know about pandas, but scikit-learn (which I guess you are going to use next) model_selection.train_test_split includes such a stratify option:
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(X, y,
stratify=y,
test_size=0.3)