How to create categorical variable based on a numerical variable

2 ways, use a couple loc calls to mask the rows where the conditions are met:

In [309]:
df.loc[(df['col1'] > 0) & (df['col1']<= 10), 'col2'] = 'xxx'
df.loc[(df['col1'] > 10) & (df['col1']<= 50), 'col2'] = 'yyy'
df.loc[df['col1'] > 50, 'col2'] = 'zzz'
df

Out[309]:
   col1 col2
0     1  xxx
1     1  xxx
2     4  xxx
3     5  xxx
4     6  xxx
5     6  xxx
6    30  yyy
7    20  yyy
8    80  zzz
9    90  zzz

Or use a nested np.where:

In [310]:
df['col2'] = np.where((df['col1'] > 0) & (df['col1']<= 10), 'xxx', np.where((df['col1'] > 10) & (df['col1']<= 50), 'yyy', 'zzz'))
df

Out[310]:
   col1 col2
0     1  xxx
1     1  xxx
2     4  xxx
3     5  xxx
4     6  xxx
5     6  xxx
6    30  yyy
7    20  yyy
8    80  zzz
9    90  zzz

You could use pd.cut as follows:

df['col2'] = pd.cut(df['col1'], bins=[0, 10, 50, float('Inf')], labels=['xxx', 'yyy', 'zzz'])

Output:

   col1 col2
0     1  xxx
1     1  xxx
2     4  xxx
3     5  xxx
4     6  xxx
5     6  xxx
6    30  yyy
7    20  yyy
8    80  zzz
9    90  zzz

You could first create a new column col2, and update its values based on the conditions:

df['col2'] = 'zzz'
df.loc[(df['col1'] > 0) & (df['col1'] <= 10), 'col2'] = 'xxx'
df.loc[(df['col1'] > 10) & (df['col1'] <= 50), 'col2'] = 'yyy'
print df

Output:

   col1 col2
0     1  xxx
1     1  xxx
2     4  xxx
3     5  xxx
4     6  xxx
5     6  xxx
6    30  yyy
7    20  yyy
8    80  zzz
9    90  zzz

Alternatively, you can also apply a function based on the column col1:

def func(x):
    if 0 < x <= 10:
        return 'xxx'
    elif 10 < x <= 50:
        return 'yyy'
    return 'zzz'

df['col2'] = df['col1'].apply(func)

and this will result in the same output.

The apply approach should be preferred in this case as it is much faster:

%timeit run() # packaged to run the first approach
# 100 loops, best of 3: 3.28 ms per loop
%timeit df['col2'] = df['col1'].apply(func)
# 10000 loops, best of 3: 187 µs per loop

However, when the size of the DataFrame is large, the built-in vectorized operations (i.e. with the masking approach) might be faster.