繁体   English   中英

如何用密度图识别离群值

[英]How to identify outliers with density plot

我正在尝试通过密度图识别异常值。 我目前正在使用seaborn库来绘制我的数据。 我将如何识别异常值? 我一直在考虑用stats库实现Z分数,这是不能在密度图中完成的唯一方法吗?

内核密度估计是通过给定数据对假设概率密度函数(pdf)的估计。 现在,我们有一个问题:哪些数据点应视为异常值。 离群值是罕见的数据点,即pdf极低的那些点。 我们不知道pdf,但知道它的估计。 因此,我们可以使用此估计来识别异常值。

因此,基本思路是:1)计算所有数据点的核密度估计; 2)找到这些点,其估计值低于某个预定义的阈值。 后者将是异常值。

让我们编写一些代码来说明这一点。

import numpy as np
# import seaborn as sns # you probably can use seaborn to get pdf-estimation values, I would use scikit-learn package for this.
from matplotlib import pyplot as plt
from sklearn.neighbors import KernelDensity

# 100 normally distributed data points and approximately 10 outliers in the end of the array.
data = np.r_[np.random.randn(100), np.random.rand(10)*100][:, np.newaxis]

# you an use kernel='gaussian' instead
kde = KernelDensity(kernel='tophat', bandwidth=0.75).fit(data)

yvals = kde.score_samples(data)  # yvals are logs of pdf-values
yvals[np.isinf(yvals)] = np.nan # some values are -inf, set them to nan

# approx. 10 percent of smallest pdf-values: lets treat them as outliers 
outlier_inds = np.where(yvals < np.percentile(yvals, 10))[0]
print(outlier_inds)
non_outlier_inds = np.where(yvals >= np.percentile(yvals, 10))[0]
print(non_outlier_inds)

[ 33  49 100 101 102 103 105 106 107 108 109]
[  0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15  16  17
  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  34  35  36
  37  38  39  40  41  42  43  44  45  46  47  48  50  51  52  53  54  55
  56  57  58  59  60  61  62  63  64  65  66  67  68  69  70  71  72  73
  74  75  76  77  78  79  80  81  82  83  84  85  86  87  88  89  90  91
  92  93  94  95  96  97  98  99 104]

# I applied log to data points because we need to visualize small (0,1) and large (up to 100) values on the same plot.
plt.plot(non_outlier_inds, np.log(data[non_outlier_inds]), 'ro',
         outlier_inds, np.log(data[outlier_inds]), 'bo')
plt.gca().set_xlabel('Index')
plt.gca().set_ylabel('log(data)')
plt.show()

在此处输入图片说明

暂无
暂无

声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.

 
粤ICP备18138465号  © 2020-2024 STACKOOM.COM