mstats_plot
Module Group¶
src/stats1
Project Stage ID¶
42
Purpose¶
The purpose of this module is to provide the user with a graphic visualisation of the statistical difference between two samples
Module Files¶
Here are the locations of the relevant files associated with the module
module information
/src/stats/mstats_plot.json
module activation functions
/src/stats/mstats_plot.py
Requirements¶
Module import information
from mllibs.nlpi import nlpi
import pandas as pd
import numpy as np
from collections import OrderedDict
import warnings; warnings.filterwarnings('ignore')
from mllibs.nlpm import parse_json
import pkg_resources
import json
import seaborn as sns
import matplotlib.pyplot as plt
import plotly.express as px
Selection¶
Activation functions need to be assigned a unique label. Here's the process of label & activation function selection
# select activation function
def sel(self,args:dict):
self.args = args
select = args['pred_task']
self.data_name = args['data_name']
self.subset = args['subset']
if(select == 'dp_hist'):
self.dp_hist(args)
if(select == 'dp_kde'):
self.dp_kde(args)
if(select == 'dp_bootstrap'):
self.dp_bootstrap(args)
if(select == 'dp_wildbootstrap'):
self.dp_wildbootstrap(args)
Activation Functions¶
Here you will find the relevant activation functions available in mstats_plot
dp_hist¶
data: [list,list] targ:None
To visualise the univariate distribution variation between two samples, we can utilise the histogram distributions. Plotly express offers the ability to compare boxplot statistics for both datasets as well
code:
# plot Histogram of Two Samples (use plotly express)
# which don't necessarily have the same sample size
def dp_hist(self,args:dict):
sample1 = args['data'][0]
sample2 = args['data'][1]
data1 = pd.DataFrame(sample1,columns=['data'])
data1['sample'] = 'one'
data2 = pd.DataFrame(sample2,columns=['data'])
data2['sample'] = 'two'
names = ['one','two']
combined = pd.concat([data1,data2])
means_data = combined.groupby(by='sample').agg('mean')
means = list(means_data['data'])
floc = [0.55,0.65]
fig = px.histogram(combined,x='data',color='sample',
marginal="box",
template='plotly_white',nbins=args['nbins'],
color_discrete_sequence=self.default_colors[0],
title='Comparing univariate distributions')
fig.update_traces(opacity=0.8)
fig.update_layout(barmode='group') # ['stack', 'group', 'overlay', 'relative']
fig.update_layout(height=350,width=700)
# fig.update_traces(marker_line_width=1,marker_line_color="white") # outline
fig.show()
sample request:
sample1 = list(np.random.exponential(scale=1, size=1000))
sample2 = list(np.random.exponential(scale=1, size=1000))
interpreter.store_data({'distribution_A':sample1,
'distribution_B':sample2})
# request
req = "compare the histograms of two samples distribution_B and distribution_A nbins 50"
# execution of request
interpreter[req]
dp_kde¶
data: [list,list] targ:None
To visualise the univariate distribution variation between two samples, we can also utilise a kernel density representation of the distributions. Seaborn offers a way to visualise this estimation in a static figure format.
code:
# plot Kernel Density Plot of Two Samples
def dp_kde(self,args:dict):
sample1 = args['data'][0]
sample2 = args['data'][1]
names = ['Sample 1','Sample 2']
fig,ax = plt.subplots(1,1,figsize=(7,3.5))
# Create a kernel density plot
sns.kdeplot(data=[sample1, sample2],palette=self.default_colors[1],ax=ax,fill=True)
sns.histplot(data=[sample1, sample2],palette=self.default_colors[1],ax=ax,alpha=0.01,stat='density',edgecolor=(0, 0, 0, 0.01))
plt.legend(names)
plt.xlabel('Values')
plt.ylabel('Density')
# plt.title('Distribution of Two Samples')
plt.title('Distribution of Two Samples', loc='left', pad=10, fontdict={'horizontalalignment': 'left'})
sns.despine(left=True)
plt.tight_layout()
plt.show()
sample request:
sample1 = list(np.random.exponential(scale=1, size=1000))
sample2 = list(np.random.exponential(scale=1, size=1000))
interpreter.store_data({'distribution_A':sample1,
'distribution_B':sample2})
# request
req = "compare kde plot of two samples distribution_B distribution_A"
# execution of request
interpreter[req]
dp_bootstrap¶
data: [list,list] targ: [nbins,nsamples]
In this method, two samples are resampled and multiple bootstrap samples are generated. Each bootstrap sample has the same size as the original sample & the mean of the distribution is stored and plotted
code:
# plot Bootstrap Histogram Distribution
def dp_bootstrap(self,args:dict):
pre = {'nsamples':100}
sample1 = np.array(args['data'][0])
sample2 = np.array(args['data'][1])
# Number of bootstrap samples
num_bootstrap_samples = self.sfp(args,pre,'nsamples')
# Perform bootstrap sampling and compute test statistic for each sample
data = {'one':[],'two':[]}
for i in range(num_bootstrap_samples):
# Resample with replacement
bootstrap_sample1 = np.random.choice(sample1, size=len(sample1), replace=True)
bootstrap_sample2 = np.random.choice(sample2, size=len(sample2), replace=True)
# Compute difference in CTR for bootstrap sample
data['one'].append(np.mean(bootstrap_sample1))
data['two'].append(np.mean(bootstrap_sample2))
fig = px.histogram(data,x=['one','two'],
marginal="box",
template='plotly_white',nbins=args['nbins'],
color_discrete_sequence=self.default_colors[0],
title='Comparing Bootstrap distributions')
fig.update_traces(opacity=0.8)
fig.update_layout(barmode='group') # ['stack', 'group', 'overlay', 'relative']
fig.update_layout(height=350,width=700)
fig.show()
sample request:
sample1 = list(np.random.exponential(scale=1, size=1000))
sample2 = list(np.random.exponential(scale=1, size=1000))
interpreter.store_data({'distribution_A':sample1,
'distribution_B':sample2})
req = "create bootstrap samples for two dataset distribution_B distribution_A nbins: 50"
interpreter[req]
dp_wildbootstrap¶
data: [list,list] targ: [nbins,nsamples]
This method is useful when dealing with heteroscedastic data or data with dependence structures. It involves resampling the residuals from a model fitted to the original data, rather than resampling the original data itself. The A/B test is performed on each bootstrap sample of residuals, and the test statistic values are recorded.
code:
# plot Wild Bootstrap Histogram Distribution
# Wild Bootstrap: This method is useful when dealing with heteroscedastic data or data with dependence structures.
# It involves resampling the residuals from a model fitted to the original data, rather than resampling the original data itself.
# The A/B test is performed on each bootstrap sample of residuals, and the test statistic values are recorded.
def dp_wildbootstrap(self,args:dict):
pre = {'nsamples':100}
sample1 = np.array(args['data'][0])
sample2 = np.array(args['data'][1])
# Number of bootstrap samples
num_bootstrap_samples = self.sfpne(args,pre,'nsamples')
# Function to estimate parameter
def estimate_parameter(data):
return np.mean(data)
# Perform Wild Bootstrap
boot1 = IIDBootstrap(sample1)
boot_estimates1 = boot1.apply(estimate_parameter, num_bootstrap_samples)
boot2 = IIDBootstrap(sample2)
boot_estimates2 = boot2.apply(estimate_parameter, num_bootstrap_samples)
data = {'one':boot_estimates1[:,0],'two':boot_estimates2[:,0]}
fig = px.histogram(data,x=['one','two'],
marginal="box",
template='plotly_white',nbins=args['nbins'],
color_discrete_sequence=self.default_colors[0],
title='Comparing Wild Bootstrap distributions')
fig.update_traces(opacity=0.8)
fig.update_layout(barmode='group') # ['stack', 'group', 'overlay', 'relative']
fig.update_layout(height=350,width=700)
fig.show()
sample request:
sample1 = list(np.random.exponential(scale=1, size=1000))
sample2 = list(np.random.exponential(scale=1, size=1000))
interpreter.store_data({'distribution_A':sample1,
'distribution_B':sample2})
req = "create wild bootstrap samples for two dataset distribution_B distribution_A nbins: 50"
interpreter[req]