Seaborn violinplots：如何获得小提琴边缘的线条路径？

Posted 2023-03-27

【中文标题】Seaborn violinplots：如何获得小提琴边缘的线条路径？

【英文标题】：Seaborn violinplots: how to I obtain line paths for violin edges?

【发布时间】：2022-01-16 12:58:47

【问题描述】：

我正在使用 Seaborn 绘制两个小提琴图，我想获得小提琴边缘的 X 坐标，这样我可以从另一个中减去一个来找到两个 KDE/分布之间的差异。我怀疑它与 matplotlib.collections.PolyCollection 对象的属性有关，但我在浏览文档时遇到了困难 - 所以我很抱歉我没有太多代码要附加......但我包括我现在的小提琴以防万一

import pandas as pd
import seaborn as sns

observed_results = [11.10283128625, 6.031906445000001, 4.625099850384, 4.371541683749999, 4.188776438315, 4.169933187839999, 4.147271982216, 3.137545605726, 2.727390606468, 2.706991071933, 2.483074510875, 2.470624399684, 2.45608460474, 2.413902898276, 2.390530982763, 2.347653087613, 2.3049660823, 2.173520711313, 2.114398085409, 2.072213409552, 1.96126510972, 1.768724290017, 1.722913211104, 1.715972042575, 1.71293343376, 1.686909025847, 1.546933962564, 1.520621928225, 1.50428319008, 1.4944074417, 1.409957657136, 1.40292975245, 1.3157577856, 1.3078804375, 1.2974806016, 1.288403682732, 1.236493409437, 1.225900768752, 1.222094652926, 1.202655483344, 1.109818003441, 1.108678687017, 1.103788138352, 1.066656041116, 0.9799193812, 0.9729697610879998, 0.9532061536159999, 0.908066599737, 0.8847958337999999, 0.8698585519490001, 0.859714675264, 0.8422146736200001, 0.8229930509580001, 0.79569571686, 0.79170962662, 0.7855221024799999, 0.775488805524, 0.7690100510069999, 0.765153773568, 0.677797640336, 0.62878587992, 0.6278724034559998, 0.619961861949, 0.555740507912, 0.5458990340579999, 0.495263271752, 0.4744517001510001, 0.453783299787, 0.426952628919, 0.4079525625, 0.4030645275, 0.401266962432, 0.3807181439999999, 0.3313936548, 0.2707379496719999, 0.256952683998, 0.248430471184, 0.242090703552, 0.235644786034, 0.214602373804, 0.199521746592, 0.1951300125, 0.173116351962, 0.172562334309, 0.156660445172, 0.145336979139, 0.1190036898, 0.114659525376, 0.094888354288, 0.06696725615, 0.0305541639, 0.023464003706, 0.021922131125]
observed_labels = ["Observed" for n in observed_results]

expected_results = [4.5217885770490405, 3.828641396489095, 3.4231762883809305, 3.1354942159291497, 2.91235066461494, 2.7300291078209855, 2.575878427993727, 2.4423470353692043, 2.324563999712821, 2.2192034840549946, 2.12389330425067, 2.03688192726104, 1.9568392195875037, 1.8827312474337816, 1.8137383759468302, 1.749199854809259, 1.6885752329928243, 1.6314168191528757, 1.5773495978825998, 1.5260563034950494, 1.4772661393256175, 1.4307461236907244, 1.3862943611198906, 1.3437347467010947, 1.3029127521808397, 1.2636920390275583, 1.2259517110447113, 1.1895840668738362, 1.1544927470625663, 1.120591195386885, 1.087801372563894, 1.0560526742493137, 1.02528101558256, 0.995428052432879, 0.9664405155596267, 0.9382696385929304, 0.9108706644048158, 0.8842024173226546, 0.858226930919394, 0.832909122935104, 0.8082165103447325, 0.7841189587656721, 0.760588461355478, 0.7375989431307791, 0.7151260872787205, 0.6931471805599453, 0.6716409753389818, 0.6505875661411494, 0.6299682789384137, 0.6097655716208943, 0.5899629443247145, 0.570544858467613, 0.5514966634969185, 0.5328045304847661, 0.5144553918165694, 0.496436886313891, 0.4787373092144902, 0.46134556650262093, 0.44425113314332093, 0.4274440148269396, 0.4109147128757291, 0.39465419200394874, 0.37865385065750773, 0.3629054936893685, 0.34740130715340317, 0.3321338350226148, 0.31709595765807425, 0.3022808718729337, 0.2876820724517809, 0.2732933349996814, 0.2591087000077249, 0.2451224580329851, 0.2313291359006492, 0.21772348384487053, 0.20430046351272993, 0.19105523676270922, 0.17798315519535654, 0.16507975035944858, 0.1523407245820189, 0.13976194237515874, 0.1273394223766015, 0.11506932978478723, 0.10294796925244237, 0.09097177820572676, 0.07913732055872386, 0.06744128079553265, 0.055880458394456614, 0.04445176257083381, 0.03315220731690051, 0.02197890671877523, 0.010929070532190317, -0.0]
expected_labels = ["Expected" for n in expected_results]

all_results = observed_results + expected_results
all_labels = observed_labels + expected_labels
df_longform = r"Z-score": all_results, 'Condition': all_labels
df_longform = pd.DataFrame(data=df_longform)

ax = sns.violinplot(x='Condition', y=r"Z-score", data=df_longform, inner=None,
                        scale='area', bw=0.3, width=0.8, saturation=1,
                        linewidth=1, cut=0)
print(ax.collections)
plt.show()

如果有关此问题的任何内容不清楚，或者我忘记在此处提供任何其他内容，请告诉我。我这辈子都想不通，所以任何建议都非常感谢 -

【问题讨论】：

通过 scipy 的 gaussian_kde 函数计算 kde 曲线可能会更容易，然后从那里寻找交点。参见例如Find non overlapping area between two kde plots in python。 violinplot 是主要的可视化工具，并不意味着精确计算。此外，您可能希望得到交叉点的近似值，而不是差值。

【参考方案1】：

以下是两者之间的比较：

calculating 并通过 kde 图及其交集可视化这两个分布 根据 kde 曲线模拟小提琴图 boxenplot，这可能是比较分布的另一种有价值的方式。

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

observed_results = [11.10283128625, 6.031906445000001, 4.625099850384, 4.371541683749999, 4.188776438315, 4.169933187839999, 4.147271982216, 3.137545605726, 2.727390606468, 2.706991071933, 2.483074510875, 2.470624399684, 2.45608460474, 2.413902898276, 2.390530982763, 2.347653087613, 2.3049660823, 2.173520711313, 2.114398085409, 2.072213409552, 1.96126510972, 1.768724290017, 1.722913211104, 1.715972042575, 1.71293343376, 1.686909025847, 1.546933962564, 1.520621928225, 1.50428319008, 1.4944074417, 1.409957657136, 1.40292975245, 1.3157577856, 1.3078804375, 1.2974806016, 1.288403682732, 1.236493409437, 1.225900768752, 1.222094652926, 1.202655483344, 1.109818003441, 1.108678687017, 1.103788138352, 1.066656041116, 0.9799193812, 0.9729697610879998, 0.9532061536159999, 0.908066599737, 0.8847958337999999, 0.8698585519490001, 0.859714675264, 0.8422146736200001, 0.8229930509580001, 0.79569571686, 0.79170962662, 0.7855221024799999, 0.775488805524, 0.7690100510069999, 0.765153773568, 0.677797640336, 0.62878587992, 0.6278724034559998, 0.619961861949, 0.555740507912, 0.5458990340579999, 0.495263271752, 0.4744517001510001, 0.453783299787, 0.426952628919, 0.4079525625, 0.4030645275, 0.401266962432, 0.3807181439999999, 0.3313936548, 0.2707379496719999, 0.256952683998, 0.248430471184, 0.242090703552, 0.235644786034, 0.214602373804, 0.199521746592, 0.1951300125, 0.173116351962, 0.172562334309, 0.156660445172, 0.145336979139, 0.1190036898, 0.114659525376, 0.094888354288, 0.06696725615, 0.0305541639, 0.023464003706, 0.021922131125]
expected_results = [4.5217885770490405, 3.828641396489095, 3.4231762883809305, 3.1354942159291497, 2.91235066461494, 2.7300291078209855, 2.575878427993727, 2.4423470353692043, 2.324563999712821, 2.2192034840549946, 2.12389330425067, 2.03688192726104, 1.9568392195875037, 1.8827312474337816, 1.8137383759468302, 1.749199854809259, 1.6885752329928243, 1.6314168191528757, 1.5773495978825998, 1.5260563034950494, 1.4772661393256175, 1.4307461236907244, 1.3862943611198906, 1.3437347467010947, 1.3029127521808397, 1.2636920390275583, 1.2259517110447113, 1.1895840668738362, 1.1544927470625663, 1.120591195386885, 1.087801372563894, 1.0560526742493137, 1.02528101558256, 0.995428052432879, 0.9664405155596267, 0.9382696385929304, 0.9108706644048158, 0.8842024173226546, 0.858226930919394, 0.832909122935104, 0.8082165103447325, 0.7841189587656721, 0.760588461355478, 0.7375989431307791, 0.7151260872787205, 0.6931471805599453, 0.6716409753389818, 0.6505875661411494, 0.6299682789384137, 0.6097655716208943, 0.5899629443247145, 0.570544858467613, 0.5514966634969185, 0.5328045304847661, 0.5144553918165694, 0.496436886313891, 0.4787373092144902, 0.46134556650262093, 0.44425113314332093, 0.4274440148269396, 0.4109147128757291, 0.39465419200394874, 0.37865385065750773, 0.3629054936893685, 0.34740130715340317, 0.3321338350226148, 0.31709595765807425, 0.3022808718729337, 0.2876820724517809, 0.2732933349996814, 0.2591087000077249, 0.2451224580329851, 0.2313291359006492, 0.21772348384487053, 0.20430046351272993, 0.19105523676270922, 0.17798315519535654, 0.16507975035944858, 0.1523407245820189, 0.13976194237515874, 0.1273394223766015, 0.11506932978478723, 0.10294796925244237, 0.09097177820572676, 0.07913732055872386, 0.06744128079553265, 0.055880458394456614, 0.04445176257083381, 0.03315220731690051, 0.02197890671877523, 0.010929070532190317, -0.0]
x0 = observed_results
x1 = expected_results

kde0 = gaussian_kde(x0, bw_method=0.3)
kde1 = gaussian_kde(x1, bw_method=0.3)

xmin = min(min(x0), min(x1))
xmax = max(max(x0), max(x1))
dx = 0.2 * (xmax - xmin)  # add a 20% margin, as the kde is wider than the data
xmin -= dx
xmax += dx

x = np.linspace(xmin, xmax, 500)
kde0_x = kde0(x)
kde1_x = kde1(x)
inters_x = np.minimum(kde0_x, kde1_x)

fig, (ax1, ax2, ax3) = plt.subplots(ncols=3, figsize=(15, 7))

ax1.plot(x, kde0_x, color='b', label='Observed')
ax1.fill_between(x, kde0_x, 0, color='b', alpha=0.2)
ax1.plot(x, kde1_x, color='orange', label='Expected')
ax1.fill_between(x, kde1_x, 0, color='orange', alpha=0.2)
ax1.plot(x, inters_x, color='r')
ax1.fill_between(x, inters_x, 0, facecolor='none', edgecolor='r', hatch='xx', label='intersection')

area_inters_x = np.trapz(inters_x, x)

handles, labels = ax1.get_legend_handles_labels()
labels[2] += f': area_inters_x * 100:.1f %'
ax1.legend(handles, labels)
ax1.set_ylim(ymin=0)
ax1.set_title("kde plot with intersection")

ax2.plot(kde0_x/2, x, color='b', label='Observed')
ax2.plot(-kde0_x/2, x, color='b')
ax2.fill_betweenx(x, -kde0_x/2, kde0_x/2, color='b', alpha=0.2)
ax2.plot(kde1_x/2, x, color='r', label='Expected')
ax2.plot(-kde1_x/2, x, color='r')
ax2.fill_betweenx(x, -kde1_x/2, kde1_x/2, color='r', alpha=0.2)
ax2.plot(inters_x/2, x, color='k')
ax2.plot(-inters_x/2, x, color='k')
ax2.fill_betweenx(x, -inters_x/2, inters_x/2, facecolor='none', edgecolor='k', hatch='oo', label='intersection')

handles, labels = ax2.get_legend_handles_labels()
labels[2] += f': area_inters_x * 100:.1f %'
ax2.legend(handles, labels)
ax2.set_title("simulated violinplot with intersection")

df_longform = pd.DataFrame(data= "Z-score": observed_results + expected_results,
                                  "Condition": ["Observed"] * len(observed_results)+["Expected"] * len(expected_results) )
sns.boxenplot(x="Condition", y="Z-score", data=df_longform, ax=ax3)
ax3.set_title("sns.boxenplot")
plt.tight_layout()
plt.show()

【讨论】：

谢谢！这正是我所要求的 =) 不知道小提琴边缘与 gaussian_kde 函数的计算完全相同。