在工作中经常会有对连续数据进行分级的工作。我们可以构造一个这样的实例:
import numpy as np from numpy.random import randomimport matplotlib.pyplot as plt %matplotlib inlinex = [a + 0.5*random() for i in range(20) for a in [1,2,2.5,3.5,4,5,6]]y = [3*random() for j in range(140)]plt.scatter(x,y,color = 'r')df = pd.DataFrame({'x':x,'y':y})
人工分级 Artificial Division
对于少量数据来说,最准确的方法当然是人工分级。
scales = [0,1.8,3.2,4.6,5.6,7]colors = ['r','g','orange','b','pink']for i in range(len(scales)-1): plt.scatter(df[(df['x']>=scales[i])&(df['x']<=scales[i+1])]['x'],df[(df['x']>=scales[i])&(df['x']<=scales[i+1])]['y'],color = colors[i])"plt.plot((1.8,1.8),(0,3.0),color = 'g')\nplt.plot((3.2,3.2),(0,3.0),color = 'g')\nplt.plot((4.6,4.6),(0,3.0),color = 'g')\nplt.plot((5.6,5.6),(0,3.0),color = 'g')"
人工分级结果,各类数据分割清晰。
等间隔分级 Equal Interval Division
数据量增大之后,难以通过肉眼观察到分界点,可以采用等间隔分级的方式进行粗暴的分级,但是通常效果不好:
x_max = max(x)x_min = min(x)scale = (x_max - x_min)/5scales = [x_min + n * scale for n in range(1,5)]scales.insert(0,x_min)scales.append(x_max)for i in range(len(scales)-1): plt.scatter(df[(df['x']>=scales[i])&(df['x']<=scales[i+1])]['x'],df[(df['x']>=scales[i])&(df['x']<=scales[i+1])]['y'],color = colors[i])
等百分比分级 Equal Percentage Division
等间隔分级常常会导致各个级别中包含的数据量悬殊,为了避免这种情况,可以将绝对间隔改为相对间隔,即采用等百分比间隔分级
x = np.array(x)scales = [np.percentile(x,20*i) for i in range(1,5)]scales.insert(0,x_min)scales.append(x_max)for i in range(len(scales)-1): plt.scatter(df[(df['x']>=scales[i])&(df['x']<=scales[i+1])]['x'],df[(df['x']>=scales[i])&(df['x']<=scales[i+1])]['y'],color = colors[i])
K均值分级 K_Means Division
分级其实是一种聚类问题,自然可以使用聚类算法,我们可以尝试用最简单的聚类算法K均值聚类来进行分级实验:
from sklearn.cluster import KMeansx = x.reshape(-1,1)km = KMeans(n_clusters=5)km.fit(x)km.labels_
output:array([3, 1, 1, 2, 2, 4, 0, 3, 1, 1, 2, 2, 4, 0, 3, 1, 1, 2, 2, 4, 0, 3, 1, 1, 2, 2, 4, 0, 3, 1, 1, 2, 2, 4, 0, 3, 1, 1, 2, 2, 4, 0, 3, 1, 1, 2, 2, 4, 0, 3, 1, 1, 2, 2, 4, 0, 3, 1, 1, 2, 2, 4, 0, 3, 1, 1, 2, 2, 4, 0, 3, 1, 1, 2, 2, 4, 0, 3, 1, 1, 2, 2, 4, 0, 3, 1, 1, 2, 2, 4, 0, 3, 1, 1, 2, 2, 4, 0, 3, 1, 1, 2, 2, 4, 0, 3, 1, 1, 2, 2, 4, 0, 3, 1, 1, 2, 2, 4, 0, 3, 1, 1, 2, 2, 4, 0, 3, 1, 1, 2, 2, 4, 0, 3, 1, 1, 2, 2, 4, 0])
import pandas as pd # x = np.squeeze(x)df['l'] = km.labels_colors = ['r','g','orange','b','pink']for i in range(5): plt.scatter(df[df['l']==i]['x'],df[df['l']==i]['y'],color = colors[i])
如上图所示,K均值分级的效果堪比人工分级。
对机器学习和算法感兴趣的朋友可以加群交流:
????