使用 scipy gaussian_kde 和 seaborn kdeplot 时的 KDE 渲染差异

Diffrence KDE rendering when using scipy gaussian_kde and seaborn kdeplot

至于 documentation tell, the seaborn kdeplot work by utilizing the scipy.stats.gaussian_kde

但是,尽管使用相同的 bandwidth 大小,但在使用 seaborngaussian_kde 绘图时,我得到了两个不同的分布。

在上图中,如果数据直接输入gaussian_kde,左边是分布。然而,如果数据输入 seaborn kdeplot.

,则正确的绘图是分布

此外,给定边界的曲线下面积在这两种绘制分布的方法之间并不相似。

auc using gaussian_kde : 47.7 and auc using via seaborn : 49.5

我可以知道是什么导致了这种差异吗?有没有一种方法可以使输出标准化,而不管使用什么方法(例如,seaborngaussian_kde

重现上述 plotauc 的代码如下。

import seaborn as sns
import matplotlib.pyplot as plt
from scipy.stats import gaussian_kde


time_window_order = ['272', '268', '264', '260', '256', '252', '248', '244', '240']
order_dict = {k: i for i, k in enumerate ( time_window_order )}
df = pd.DataFrame ( {'time_window': ['268', '268', '268', '264', '252', '252', '252', '240',
                                     '256', '256', '256', '256', '252', '252', '252', '240'],
                     'seq_no': ['a', 'a', 'a', 'a', 'a', 'a', 'a', 'a',
                                'b', 'b', 'b', 'b', 'b', 'b', 'b', 'b']} )
df ['centre_point'] = df ['time_window'].map ( order_dict )
filter_band = df ["seq_no"].isin ( ['a'] )
df = df [filter_band].reset_index ( drop=True )
auc_x_min, auc_x_max = 0, 4
bandwith=0.5
########################

plt.subplots(1, 2)
# make the first plot
plt.subplot(1, 2, 1)
kde0 = gaussian_kde ( df ['centre_point'], bw_method=bandwith )
xmin, xmax = -3, 12
x_1 = np.linspace ( xmin, xmax, 500 )
kde0_x = kde0 ( x_1 )
sel_region_x = x_1 [(x_1 > auc_x_min) * (x_1 < auc_x_max)]
sel_region_y = kde0_x [(x_1 > auc_x_min) * (x_1 < auc_x_max)]
auc_bond_1 = np.trapz ( sel_region_y, sel_region_x )
area_whole = np.trapz ( kde0_x, x_1 )
plt.plot ( x_1, kde0_x, color='b', label='KDE' )
plt.ylim(bottom=0)
plt.title(f'Direct gaussian_kde with bw {bandwith}')
plt.fill_between ( sel_region_x, sel_region_y, 0, facecolor='none', edgecolor='r', hatch='xx',
                   label='intersection' )

# make second plot
plt.subplot(1, 2, 2)

g = sns.kdeplot ( data=df, x="centre_point", bw_adjust=bandwith )
c = g.get_lines () [0].get_data ()
x_val = c [0]
kde0_x = c [1]
idx = (x_val> auc_x_min) * (x_val < auc_x_max)
sel_region_x = x_val [idx]
sel_region_y = kde0_x [idx]
auc_bond_2 = np.trapz ( sel_region_y, sel_region_x )
g.fill_between ( sel_region_x, sel_region_y, 0, facecolor='none', edgecolor='r', hatch='xx' )
plt.title(f'Via Seaborn with bw {bandwith}')
plt.tight_layout()
plt.show()

# show much the area differ between these two plotting
print ( f'auc using gaussian_kde : {auc_bond_1 * 100:.1f} and auc using via seaborn : {auc_bond_2 * 100:.1f}' )
x=1

编辑

的基础上,对这两行进行改动

kde0 = gaussian_kde ( df ['centre_point'], bw_method='scott' )

g = sns.kdeplot ( data=df, x="centre_point", bw_adjust=1 ) # Seaborn by default use the scott method to determine the bw size

return

从视觉上看,这两个图看起来是一样的。

但是,图表之间的auc仍然是return两个不同的值

auc using gaussian_kde : 45.1 and auc using via seaborn : 44.6

您正在这样呼叫 scipy:

kde0 = gaussian_kde ( df ['centre_point'], bw_method=bandwith )

像这样的 seaborn

g = sns.kdeplot ( data=df, x="centre_point", bw_adjust=bandwith )

但是 kdeplot docs 告诉我们 bw_adjust

Factor that multiplicatively scales the value chosen using bw_method. Increasing will make the curve smoother. See Notes.

而 kdeplot 也有一个 bw_method 参数是

Method for determining the smoothing bandwidth to use; passed to scipy.stats.gaussian_kde.

因此,如果您想使两个库的结果相等,则需要确保使用了正确的参数。