Fill_NaNs_Numeric.py

valueCounts = {}
def CountAll():
    global all_columns, nanCounts, valueCounts
    all_columns = list(df)
    nanCounts = df.isnull().sum()
    for x in all_columns:
        valueCounts[x] = df[x].value_counts()

"""Random but proportional replacement(RBPR) of numeric"""
def Fill_NaNs_Numeric(col):

    mini = df[col].min()
    maxi = df[col].max()
    """Selecting ONLY non-NaNs."""
    temp = df[df[col].notnull()][col] # type --> pd.Series

    """Any continuous data is 'always' divided into 45 bins (Hard-Coded)."""
    bin_size = 45
    bins = np.linspace(mini, maxi, bin_size)

    """Filling the bins (with non-NaNs) and calculating mean of each bin."""
    non_NaNs_per_bin = []
    mean_of_bins = []

    non_NaNs_per_bin.append(len(temp[(temp <= bins[0])]))
    mean_of_bins.append(temp[(temp <= bins[0])].mean())
    for x in range(1, bin_size):
        non_NaNs_per_bin.append(len(temp[(temp <= bins[x]) & (temp > bins[x-1])]))
        mean_of_bins.append(temp[(temp <= bins[x]) & (temp > bins[x-1])].mean())

    mean_of_bins = pd.Series(mean_of_bins)
    # np.around() on  list 'proportion' may create trouble and we may get a zero-value imputed, hence,
    mean_of_bins.fillna(temp.mean(), inplace= True)
    non_NaNs_per_bin = np.array(non_NaNs_per_bin)

    """Followoing part is SAME as Fill_NaNs_Catigorical()"""

    """Calculating probability and expected value."""
    proportion = np.array(non_NaNs_per_bin) / valueCounts[col].sum() * nanCounts[col]
    proportion = np.around(proportion).astype('int')

    """Adjusting proportion."""
    diff = int(nanCounts[col] - np.sum(proportion))
    if diff > 0:
        for x in range(diff):
            idx = random.randint(0, len(proportion) - 1)
            proportion[idx] =  proportion[idx] + 1
    else:
        diff = -diff
        while(diff != 0):
            idx = random.randint(0, len(proportion) - 1)
            if proportion[idx] > 0:
                proportion[idx] =  proportion[idx] - 1
                diff = diff - 1

    """Filling NaNs."""
    nan_indexes = df[df[col].isnull()].index.tolist()
    for x in range(len(proportion)):
            if proportion[x] > 0:
                random_subset = random.sample(population= nan_indexes, k= proportion[x])
                df.loc[random_subset, col] = mean_of_bins[x] # <--- Replacing with bin mean
                nan_indexes = list(set(nan_indexes) - set(random_subset))

"""-------------------------------------------------------------------------"""

Hey, @Vernal-Inertia, before anything, thanks for sharing the code and writing the article on RBPR; it's quite elucidating!
I was wondering if maybe you could explain a little further the proportion calculus? It's not totally clear for me yet.

Hello @abcorep, thank you for taking out the time to read the article. Understanding the mathematics part is actually simple. Assume there are 2 boys and 3 girls; now you have to distribute 10 chocolates to them? One answer is distributing 2 chocolates to each. But with a different perspective we can also have another solution - where in we distribute chocolates in proportion to the gender:

(no. of girls)/(total children) = 3/5 --> this the ratio of all girls to total children.
So how many chocolates go to only girls? --> (3/5)*10 = 6
And how many chocolates go to only boys? --> (2/5)*10 = 4

Thus 10(number of chocolates) has been split into 6:4 which is actually 3:2, that is ratio of girls to boys. Hope this resolves your doubt.