gaussian_encodings.py

import numpy as np
from scipy.spatial.distance import pdist, squareform

# function that converts categorical variable 
# into a one-hot encoding
def one_hot_encoding(x):
    n = len(x)    
    min_category = np.min(x)
    max_category = np.max(x)
    num_categories = max_category - min_category + 1
    output = np.zeros((n, num_categories))
    for i,x_i in enumerate(x):
        output[i, x_i - min_category] = 1
    return output
    
    
# function that converts categorical variable
# into a random gaussian encoding of length 
def gaussian_encoding(x,gaussian_encodings=None):
    n = len(x)    
    min_category = np.min(x)
    max_category = np.max(x)
    num_categories = max_category - min_category + 1
    encoding_dimension = int(np.log(num_categories))
    output = np.zeros((n, encoding_dimension))
    if gaussian_encodings is None:
        gaussian_encodings = dict()
    for i,x_i in enumerate(x):
        if x_i not in gaussian_encodings:
            gaussian_encodings[x_i] = np.random.normal(loc=0,scale=1,size=encoding_dimension)
        output[i,:] = gaussian_encodings[x_i]
    return output, gaussian_encodings
    
    
# test data
n = 1000
num_categories = 100
x = np.random.randint(low=0,high=num_categories,size=n)

# get the encodings
x_one_hot = one_hot_encoding(x)
x_gaussian, gaussian_encodings = gaussian_encoding(x)