Pandas dataframe, select n 基于唯一值数量的随机行

Question

I'm working on a text classification problem that trains well but my categories are quite imbalanced, hindering results. The largest 2 categories are over 80x larger than the smallest category, so an unfair amount of the classifications go to those 2 categories. I need to select n rows (arbitrarily large) from each category. My dataset is quite large (10m rows, 1k unique categories).

Let's say the dataframe is:

data = {
    'category':['2','2','2','2','4','4','4','4','4','4','6','6','6'],
    'text':['t1','t2','t3','t4','t5','t6','t7','t8','t9','t10','t11','t12','t13']
}

df = pd.DataFrame(data)

How could I select n random rows per category?

I have tried to find some way to use np.random.choice to select n random rows but I can't find a way to grab that index for a drop by index.

The ideal output for n = 3 would be something like:

>>> df.head(9)
    category    text
0   2           t3
1   6           t11
2   6           t13
3   4           t6
4   2           t1
5   4           t9
6   4           t8
7   2           t4
8   6           t12

Answer 1

You can use sample and groupby().head() :

df.sample(frac=1).groupby('category').head(3)

Output:

   category text
4         4   t5
12        6  t13
1         2   t2
8         4   t9
9         4  t10
3         2   t4
10        6  t11
0         2   t1
11        6  t12