Relief 过滤式特征选择

给定训练集{(x1,y1),(x2,y2).....(xm,ym)} ，对每个示例xi，Relief在xi的同类样本中寻找其最近邻x_i,nh（猜中近邻），再从xi的异类样本中寻找其最近邻x_i,nm(猜错近邻)

代码如下：

#!/usr/bin/env python2
# -*- coding: utf-8 -*-
"""
Created on Wed Feb 28 20:16:09 2018

@author: jzc
"""
import numpy as np
import csv
from random import randrange
from sklearn import preprocessing
#抽样次数
m=8
def Compute_Distance_Discrete(diff_distance):
    # 使用欧几里得距离计算最近邻
    counter = np.power(diff_distance,2)
    counter = np.sum(counter)
    counter = np.sqrt(counter)
    return counter
def loadcsv(filename):
    """载入文件"""
    lines = csv.reader(open(filename,'r'))
    data = list(lines)
    for i in range(1,len(data)):
        data[i] = [float(x) for x in data[i]]
    result = np.array(data[1:])
    features = result[:,1:-1]
    labels = result[:,-1]
    return features,labels
def Relief(features,labels):
    #初始化
    (n_samples,n_features)=np.shape(features)
    distance = np.zeros((n_samples,n_samples))
    weights = np.zeros(n_features)
    nearHit= list()
    nearMiss= list()
    distance_sort=list()
    """寻找每个样本的距离"""
    for i in range(0,n_samples):
        for j in range(0,n_samples):
            diff_distance = features[i]-features[j]
            if i==j:
                distance[i,j]=99999
            else:
                distance[i,j] = Compute_Distance_Discrete(diff_distance)

    for i in range(0,m):
        one_sample = randrange(0,n_samples,1) #随机选择一个样本
        one_feature = features[one_sample]
        for index in range(n_samples):
            distance_sort.append([distance[one_sample,index],index,labels[index]])
        #从小到大排序
        distance_sort.sort(key = lambda x:x[0])
        """寻找样本的猜错近邻和猜中近邻"""
        for index in range(n_samples):
            if nearHit ==[] and distance_sort[index][2]==labels[one_sample]:
                nearHit = features[distance_sort[index][1]]
            elif nearMiss==[] and distance_sort[index][2]!=labels[one_sample]:
                nearMiss = features[distance_sort[index][1]]
            elif nearHit!=[] and nearMiss!=[]:
                break;
            else:
                continue;
        sum_nh = list()
        sum_nm =list()
        # 若属性j离散，Xaj==Xbj 则diff的值为0；否则为1
        for k in range(len(one_feature[:-2])):
            if one_feature[k] != nearHit[k]:
                sum_nh.append(1)
            else:
                sum_nh.append(0)
            if one_feature[k] != nearMiss[k]:
                sum_nm.append(1)
            else:
                sum_nm.append(0)
        #print sum_nh,sum_nm
        #print one_feature[-2:]-nearHit[-2:]
        """若为属性j为连续， diff(Xaj-Xbj)=|Xaj-Xbj| 并且Xaj,Xbj要归一化到[0,1]区间"""
        weights[-2:] = weights[-2:]-np.power(one_feature[-2:]-nearHit[-2:],2)
        +np.power(one_feature[-2:]-nearMiss[-2:],2)
        weights[:-2] = weights[:-2]-np.power(sum_nh,2)+np.power(sum_nm,2)
        #print weights/n_samples
    return weights/n_samples
filename = '/Users/jzc/DeepLearning(7.8-)/data/watermelon3_0.csv'
features,labels = loadcsv(filename)
#features[-2:] = preprocessing.normalize(features[-2:],norm='l2')
#print features
for x in range(1,10):
    result = Relief(features,labels)
    print result
#print features[0],labels[0]