Replace NaN in a dataframe with random values

Question

I have a data frame (data_train) with NaN values, A sample is given below:

republican                n                          y   
republican                n                          NaN   
democrat                 NaN                         n
democrat                  n                          y   

I want to replace all the NaN with some random values like .

republican                n                           y   
republican                n                          rnd2
democrat                 rnd1                         n
democrat                  n                           y   

How do I do it.

I tried the following, but had no luck:

df_rand = pd.DataFrame(np.random.randn(data_train.shape[0],data_train.shape[1]))
data_train[pd.isnull(data_train)] = dfrand[pd.isnull(data_train)]

when I do the above with a dataframe with random numerical data the above script works fine.

Solution 1

You can use the pandas update command, this way:

1) Generate a random DataFrame with the same columns and index as the original one:

import numpy as np; import pandas as pd
M = len(df.index)
N = len(df.columns)
ran = pd.DataFrame(np.random.randn(M,N), columns=df.columns, index=df.index)

2) Then use update, so that the NaN values in df will be replaced by the generated random values

df.update(ran)

---

In the above example I used values from a standard normal, but you can also use values randomly picked from the original DataFrame:

import numpy as np; import pandas as pd

M = len(df.index)
N = len(df.columns)

val = np.ravel(df.values)
val = val[~np.isnan(val)]
val = np.random.choice(val, size=(M,N))
ran = pd.DataFrame(val, columns=df.columns, index=df.index)

df.update(ran)

Solution 2

Well, if you use fillna to fill the NaN, a random generator works only once and will fill all N/As with the same number.

So, make sure that a random number is generated and used each time. For a dataframe like this :

          Date         A       B
0   2015-01-01       NaN     NaN
1   2015-01-02       NaN     NaN
2   2015-01-03       NaN     NaN
3   2015-01-04       NaN     NaN
4   2015-01-05       NaN     NaN
5   2015-01-06       NaN     NaN
6   2015-01-07       NaN     NaN
7   2015-01-08       NaN     NaN
8   2015-01-09       NaN     NaN
9   2015-01-10       NaN     NaN
10  2015-01-11       NaN     NaN
11  2015-01-12       NaN     NaN
12  2015-01-13       NaN     NaN
13  2015-01-14       NaN     NaN
14  2015-01-15       NaN     NaN
15  2015-01-16       NaN     NaN

I used the following code to fill up the NaNs in column A:

import random
x['A'] = x['A'].apply(lambda v: random.random() * 1000)

Which will give us something like:

          Date           A       B
0   2015-01-01   96.538211     NaN
1   2015-01-02  404.683392     NaN
2   2015-01-03  849.614253     NaN
3   2015-01-04  590.030660     NaN
4   2015-01-05  203.167519     NaN
5   2015-01-06  980.508258     NaN
6   2015-01-07  221.088002     NaN
7   2015-01-08  285.013762     NaN

Solution 3

If you want to replace all NaNs from the DF with random values from a list, you can do something like this:

import numpy as np

df.applymap(lambda l: l if not np.isnan(l) else np.random.choice([1, 3]))

Solution 4

If you want to replace NaN in your column with hot deck technique, I can propose way like this :

def hot_deck(dataframe) :
    dataframe = dataframe.fillna(0)
    for col in dataframe.columns :
        assert (dataframe[col].dtype == np.float64) | (dataframe[col].dtype == np.int64)
        liste_sample = dataframe[dataframe[col] != 0][col].unique()
        dataframe[col] = dataframe.apply(lambda row : random.choice(liste_sample) if row[col] == 0 else row[col],axis=1)
    return dataframe

After if you prefer just replace NaN with a new random value for each iteration you can do a thing like that. You've just to determine the max value of your random choices.

def hot_deck(dataframe,max_value) :
    dataframe = dataframe.fillna(0)
    for col in dataframe.columns :
        assert (dataframe[col].dtype == np.float64) | (dataframe[col].dtype == np.int64)
        liste_sample = random.sample(range(max_value),dataframe.isnull().sum())
        dataframe[col] = dataframe.apply(lambda row : random.choice(liste_sample) if row[col] == 0 else row[col],axis=1)
    return dataframe

View more solutions