For a binary classification, how can I sample data by column so that 2/3 of the data contains zeros and 1/3 contains ones?

Question

I have a large dataset containing four columns and the third column contains the binary label (a value either 0 or 1). This dataset is imbalanced - it contains much more zeros than ones. The data looks like:

3   5   0   0.4
4   5   1   0.1
5   13  0   0.5
6   10  0   0.8
7   25  1   0.3
:   :   :   :

I know that I can obtain a balanced subset containing 50% zeros and 50% ones by for example:

df_sampled = df.groupby(df.iloc[:,2]).sample(n=20000, random_state=1)

But how I can amend the one-liner given above to change the ratio of zeros and ones? For example how can I sample this data (by the third column) so that 2/3 of the data contains zeros and 1/3 contains ones?

Answer 1

This is a possible solution:

n_samples = 90000 # total number of samples

df_sampled = pd.concat(
    [group.sample(n=int(n_samples * 2 / 3)) if label == 0
     else group.sample(n=int(n_samples * 1 / 3))
     for label, group in df.groupby(df.iloc[:, 2])]
)

A similar solution would be:

n_samples = 90000 # total number of samples
ratios = [2 / 3, 1 / 3]

df_sampled = pd.concat(
    [group.sample(n=int(n_samples * ratios[label]))
     for label, group in df.groupby(df.iloc[:, 2])]
)

Here I'm basically applying a different function to different groups.