Python Implementation of: "Fast Generation of Accurate Synthetic Microdata" https://crises-deim.urv.cat/web/docs/publications/lncs/443.pdf

micordata.py

import numpy as np

def compute_mean(v):
    return sum(v)/len(v)

def subtract_mean_to_col(M, idx):
    M[:, idx] = M[:, idx] - compute_mean(M[:, idx])

def alg_two(n=None, m=None, A=None):
    #1
    # we can provide dimensions or a random matrix 
    if n is None and A is None:
        raise ValueError("At least one parameter among n and A should be passed")
    #if random matrix is not provided then generate one!
    if A is None:
        if m is None:
            m = n

        A = np.random.rand(n, m)

    else:
        (n, m) = np.shape(A)
    #2    
    subtract_mean_to_col(A, 0)
    #3
    for i in range(1, m):
        A[-i:,i] = np.linalg.solve(A[-i:,:i].T, -A[:-i,:i].T.dot(A[:-i,i]))
        subtract_mean_to_col(A,i)
    #4
    return A / np.std(A,0)


def alg_one(X):
    # 1
    A = alg_two(*X.shape)
    # 2
    C_x = np.cov(X,ddof=0, rowvar=False)
    # 3
    U = np.linalg.cholesky(C_x).T
    # 4
    X_I = A.dot(U)
    # 5
    X_col_means = np.mean(X,0)
    return X_I + X_col_means