IFLS.py

# Infinite Feature Selection.
# https://pypi.org/project/PyIFS/
# Inputs:
#   x_train: matrix T (samples) by n (number of features).
#   y_train: column vector of labels (-1,1).
#   verbose: boolean variable.

# Outputs:
#   RANKED: vector of indicies of x_train from the best to the worst feature.

import numpy as np
import math
# import stats

class InfFS:

    def __init__(self):
        h=0.1
        #print("HI")
    # Take in input the matrix e the label vector and return a matrix
    # of data for every different label.
    def takeLabel(self, x_train, y_train ):
        counter = x_train.shape[0] -1
        s_n = x_train
        s_p = x_train
        while(1):
            if( y_train[counter] == 1 ):
                s_n = np.delete(s_n, counter, axis = 0 )
            else:
                s_p = np.delete(s_p, counter, axis = 0 )
            counter = counter - 1
            if( counter == - 1 ):
                break
        return s_p, s_n

    # Function that help to define priors_corr.
    def defPriorsCorr(self,mu_s_n, mu_s_p):
        pcorr = mu_s_p
        counter = 0
        while( counter < len(pcorr) ):
            pcorr[counter] = (pcorr[counter] - mu_s_n[counter])*(pcorr[counter] - mu_s_n[counter])
            counter = counter + 1
        return pcorr

    # Function to subtract the min value of the matrix to all it's elements.
    def SubtractMin(self, corr_ij ):
        m = 10100
        for i in range(0,corr_ij.shape[0]): # Find the min.
            for j in range(0,corr_ij.shape[1]):
                if( corr_ij[i,j] < m ):
                    m = corr_ij[i,j]

        for i in range(0,corr_ij.shape[0]): # Subtract the min value.
            for j in range(0,corr_ij.shape[1]):
                corr_ij[i,j] = corr_ij[i,j] - m

        return corr_ij

    # Function to divide every element of the matrix to his maximum value.
    def DivideByMax(self,corr_ij):
        m = -1
        for i in range(0,corr_ij.shape[0]): # Find the max.s
            for j in range(0,corr_ij.shape[1]):
                if( corr_ij[i,j] > m ):
                    m = corr_ij[i,j]

        for i in range(0,corr_ij.shape[0]): # Divide by the maximum value.
            for j in range(0,corr_ij.shape[1]):
                corr_ij[i,j] = corr_ij[i,j] / m

        return corr_ij

    # Handmaded bsxfunction that take the max.
    def bsxfun(self, STD ):
        m = np.zeros( (STD.shape[0], STD.shape[0]) )
        for i in range( 0,STD.shape[0] ):
            for j in range( 0,STD.shape[0] ):
                if( STD[i] > STD[j] ):
                    m[i,j] = STD[i]
                else:
                    m[i,j] = STD[j]
        return m

    def infFS(self,x_train, y_train, alpha, supervision, verbose):
        # Start of point one.

        if supervision:
            s_p, s_n = self.takeLabel( x_train, y_train)
            mu_s_n = s_n.mean(0)
            mu_s_p = s_p.mean(0)
            priors_corr = self.defPriorsCorr(mu_s_n, mu_s_p)
            st = np.power(np.std(s_p, ddof = 1, axis = 0),2)
            st = st + np.power(np.std(s_n,ddof = 1, axis = 0),2)
            for i in range(0,len(st)):
                if st[i] == 0:
                    st[i] = 10000
            corr_ij = priors_corr
            for i in range(0,len(corr_ij)):
                corr_ij[i] = corr_ij[i] / st[i]
            corr_ij = np.dot( corr_ij.T[:,None], corr_ij[None,:])
            corr_ij = self.SubtractMin(corr_ij)
            corr_ij = self.DivideByMax(corr_ij)
        else:
            corr_ij, pval = stats.spearmanr(x_train)
            for i in range( 0,corr_ij.shape[0] ):
                for j in range( 0,corr_ij.shape[1] ):
                    if( math.isnan(corr_ij[i,j]) or corr_ij[i,j] < -1 or corr_ij[i,j] > 1 ):
                        corr_ij[i,j] = 0

        # After if.
        STD = np.std(x_train, ddof = 1, axis = 0)
        STDMatrix = self.bsxfun( STD )
        STDMatrix = self.SubtractMin(STDMatrix)
        sigma_ij = self.DivideByMax(STDMatrix)
        for i in range( 0,sigma_ij.shape[0] ):
            for j in range( 0,sigma_ij.shape[1] ):
                if( math.isnan(sigma_ij[i,j]) or sigma_ij[i,j] < -1 or sigma_ij[i,j] > 1 ):
                    sigma_ij[i,j] = 0

        # End of point one.

        # Start of the point two.
        if (verbose):
            print("2) Building the graph G = <V,E> \n");
        A =  ( alpha*corr_ij + (1-alpha)*sigma_ij );
        # End of the point two.

        # Start of the point three.
        if (verbose):
            print("3) Letting paths tend to infinite \n");

        I = np.identity( A.shape[0] )
        r = ( 0.9/ max( np.linalg.eigvals(A) ) ) # Setting the r values.
        y = I - ( r * A )
        S = np.linalg.inv( y ) - I
        # End of point three.

        # Start of point four.
        if (verbose):
            print("4) Estimating energy scores \n")

        WEIGHT = np.sum( S , axis=1 )
        # End of point four.

        # Start of point five.
        if(verbose):
            print("5) Features ranking")

        RANKED = np.argsort(WEIGHT)
        RANKED = np.flip(RANKED,0)
        RANKED = RANKED.T
        WEIGHT = WEIGHT.T
        return RANKED, WEIGHT
        # End of point five.