Identifying Outliers with Quantile Regression and Python - python

I am trying to identify outliers in a dataset using the 5th and 95th percentiles of a regression line so I'm using quantile regression in Python with statsmodel, matplotlib and pandas. Based on this answer from blokeley, I can create a scatterplot of my data and show the best fit line and the lines for the 5th and 95th percentile based on quantile regression. But how do I identify those points that fall above and below those lines and then save them out to a pandas dataframe?
My data looks like this (there are 95 values in total):
Month Year LST NDVI
0 June 1984 310.550975 0.344335
1 June 1985 310.495331 0.320504
2 June 1986 306.820900 0.369494
3 June 1987 308.945602 0.369946
4 June 1988 308.694022 0.31863
2
and the script I have so far is this:
import pandas as pd
excel = my_excel
df = pd.read_excel(excel)
df.head()
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import statsmodels.formula.api as smf
model = smf.quantreg('NDVI ~ LST',df)
quantiles = [0.05,0.95]
fits = [model.fit(q=q) for q in quantiles]
figure,axes = plt.subplots()
x = df['LST']
y = df['NDVI']
axes.scatter(x,df['NDVI'],c='green',alpha=0.3,label='data point')
fit = np.polyfit(x, y, deg=1)
axes.plot(x, fit[0] * x + fit[1], color='grey',label='best fit')
_x = np.linspace(x.min(),x.max())
for index, quantile in enumerate(quantiles):
_y = fits[index].params['LST'] * _x + fits[index].params['Intercept']
axes.plot(_x, _y, label=quantile)
title = 'LST/NDVI Jun-Aug'
plt.title(title)
axes.legend()
axes.set_xticks(np.arange(298,320,4))
axes.set_yticks(np.arange(0.25,0.5,.05))
axes.set_xlabel('LST')
axes.set_ylabel('NDVI');
And the chart I get out of that is this:
So I can definitely see data points above the 95th line and below the 5th line that I would classify as outliers, but I want to identify those in my original dataframe and maybe plot them on the cart or highlight them in some way to show them as "outliers".
I am searching on a method but coming up empty and could use some help.

You need to figure out if certain point are above the 95% quantile line or below the 5% quantile line. This you can do using the cross product, see this answer for a straightforward implementation.
In your example, you would need to combine the points above and below the quantile lines, possibly in a mask.
Here's is an example:
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import statsmodels.formula.api as smf
df = pd.DataFrame(np.random.normal(0, 1, (100, 2)))
df.columns = ['LST', 'NDVI']
model = smf.quantreg('NDVI ~ LST', df)
quantiles = [0.05, 0.95]
fits = [model.fit(q=q) for q in quantiles]
figure, axes = plt.subplots()
x = df['LST']
y = df['NDVI']
fit = np.polyfit(x, y, deg=1)
_x = np.linspace(x.min(), x.max(), num=len(y))
# fit lines
_y_005 = fits[0].params['LST'] * _x + fits[0].params['Intercept']
_y_095 = fits[1].params['LST'] * _x + fits[1].params['Intercept']
# start and end coordinates of fit lines
p = np.column_stack((x, y))
a = np.array([_x[0], _y_005[0]]) #first point of 0.05 quantile fit line
b = np.array([_x[-1], _y_005[-1]]) #last point of 0.05 quantile fit line
a_ = np.array([_x[0], _y_095[0]])
b_ = np.array([_x[-1], _y_095[-1]])
#mask based on if coordinates are above 0.95 or below 0.05 quantile fitlines using cross product
mask = lambda p, a, b, a_, b_: (np.cross(p-a, b-a) > 0) | (np.cross(p-a_, b_-a_) < 0)
mask = mask(p, a, b, a_, b_)
axes.scatter(x[mask], df['NDVI'][mask], facecolor='r', edgecolor='none', alpha=0.3, label='data point')
axes.scatter(x[~mask], df['NDVI'][~mask], facecolor='g', edgecolor='none', alpha=0.3, label='data point')
axes.plot(x, fit[0] * x + fit[1], label='best fit', c='lightgrey')
axes.plot(_x, _y_095, label=quantiles[1], c='orange')
axes.plot(_x, _y_005, label=quantiles[0], c='lightblue')
axes.legend()
axes.set_xlabel('LST')
axes.set_ylabel('NDVI')
plt.show()

Related

Recover the time shift from nympy.correlate result in Python

This is not a duplicate question since other answers only explain how to plot the cross-correlation function and do not explain how you can get the time difference.
Given a sin signal and shifted version, we should be able to get the time delay between them.
I have created a sin signal and shifted it by t_d=0.05. The following is my code and its output:
import numpy as np
import matplotlib.pyplot as plt
fs = 1000
x = np.linspace(0, 1, fs)
f = 5
t_shift = 0.05
y = np.sin(2*np.pi*f*x)
y_shifted = np.sin(2*np.pi*f*(x-t_shift))
fig, ax = plt.subplots()
ax.plot(x, y, x, y_shifted)
plt.show()
By normalizing signals and applying numpy.correlate we get the following:
y_norm = (y-y.mean())/y.std()
y_shifted_norm = (y_shifted - y_shifted.mean())/y_shifted.std()
cc = np.correlate(y_norm, y_shifted_norm, 'full')
fig, ax = plt.subplots()
ax.plot(range(len(cc)), cc)
plt.show()
Question
From the indices of cross-correlation function, how can I get t_shift=0.05?
#Sepide. It seems to me as if you are trying to maximise the correlation between the signal y and a shifted version of y_shifted. This might be accomplished using np.correlate() but it seems nontrivial indeed to recover the time shifts in the signals. In the solution below I manually shift the time series and compute the correlation coefficient using np.corrcoef. As soon as this Pearson correlation coefficient equals 1, the two signals are aligned.
import numpy as np
import matplotlib.pyplot as plt
# Setting
fs = 1000
x = np.linspace(0, 1, fs)
f = 5
t_shift = 0.05
t_step = 1/fs
# Data
y = np.sin(2*np.pi*f*x)
y_shifted = np.sin(2*np.pi*f*(x-t_shift))
# Compute correlation
MaxTimeShift = 200
CorrelationList = np.empty((MaxTimeShift,1));
CorrelationList[:] = np.NaN
# Compute correlation for various shifts
for iter in range(MaxTimeShift):
CorrelationList[iter] = np.corrcoef( y[0:801].T, y_shifted[iter:(801+iter)].T)[0,1]
# Plot 1
plt.figure(1)
plt.plot(x, y, x, y_shifted)
plt.show()
# Plot 2
plt.figure(2)
ShiftList = t_step*np.arange(MaxTimeShift)
plt.plot(ShiftList, CorrelationList)
plt.title("Correlation coefficient")
plt.show()
print("The time shift between the signals is: ", ShiftList[np.argmax(CorrelationList)])

Plotting and modeling data with lmfit - Fit doesn't match data. What am I doing wrong?

I have some data I'm trying to model with lmfit's Model.
Specifically, I'm measuring superconducting resistors. I'm trying fit the experimental data (resistance vs. temperature) to a model which incorporates the critical temperature Tc (material dependent), the resistance below Tc (nominally 0), and the resistance above Tc (structure dependent).
Here's a simplified version (with simulated data) of the code I'm using to plot my data, along with the output plot.
I'm not getting any errors but, as you can see, I'm also not getting a fit that matches my data.
What am I doing wrong? This is my first time using lmfit and Model, so I may be making a newbie mistake. I thought I was following the lmfit example but, as I said, I'm obviously doing something wrong.
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from lmfit import Model
def main():
x = np.linspace(0, 12, 50)
x_ser = pd.Series(x) # Simulated temperature data
y1 = [0] * 20
y2 = [10] * 30
y1_ser = pd.Series(y1) # Simulated resistance data below Tc
y2_ser = pd.Series(y2) # Simulated resistance data above Tc (
y_ser = y1_ser.append(y2_ser, ignore_index=True)
xcrit_model = Model(data_equation)
params = xcrit_model.make_params(y1_guess=0, y2_guess=12, xcrit_guess=9)
print('params: {}'.format(params))
result = xcrit_model.fit(y_ser, params, x=x_ser)
print(result.fit_report())
plt.plot(x_ser, y_ser, 'bo', label='simulated data')
plt.plot(x_ser, result.init_fit, 'k.', label='initial fit')
plt.plot(x_ser, result.best_fit, 'r:', label='best fit')
plt.legend()
plt.show()
def data_equation(x, y1_guess, y2_guess, xcrit_guess):
x_lt_xcrit = x[x < xcrit_guess]
x_ge_xcrit = x[x >= xcrit_guess]
y1 = [y1_guess] * x_lt_xcrit.size
y1_ser = pd.Series(data=y1)
y2 = [y2_guess] * x_ge_xcrit.size
y2_ser = pd.Series(data=y2)
y = y1_ser.append(y2_ser, ignore_index=True)
return y
if __name__ == '__main__':
main()
lmfit (and basically all similar solvers) work with continuous variables and investigate how they alter the result by making tiny changes in the parameter values and seeing how that effects this fit.
But your xcrit_guess parameter is used only as a discrete variable. If its value changes from 9.0000 to 9.00001, the fit will not change at all.
So, basically, don't do:
x_lt_xcrit = x[x < xcrit_guess]
x_ge_xcrit = x[x >= xcrit_guess]
Instead, you should use a smoother sigmoidal step function. In fact, lmfit has one of these built-in. So you might try something like this (note, there is no point in converting numpy.arrays to pandas.Series - the code will just turn these back to numpy arrays anyway):
import numpy as np
from lmfit.models import StepModel
import matplotlib.pyplot as plt
x = np.linspace(0, 12, 50)
y = 9.5*np.ones(len(x))
y[:26] = 0.0
y = y + np.random.normal(size=len(y), scale=0.0002)
xcrit_model = StepModel(form='erf')
params = xcrit_model.make_params(amplitude=4, center=5, sigma=1)
result = xcrit_model.fit(y, params, x=x)
print(result.fit_report())
plt.plot(x, y, 'bo', label='simulated data')
plt.plot(x, result.init_fit, 'k', label='initial fit')
plt.plot(x, result.best_fit, 'r:', label='best fit')
plt.legend()
plt.show()

Smoothing / noise filtering data in Python

I have table with data as follows
article price wished outcome
horse 10 10
duck 15 15
child 9 15 - 21
panda 21 21
lamb 24 22
gorilla 23 23
I want to smooth column Price to the wished Price and then put it into dataframe, so that I see the values.
Please, is there some built in library - method that smoothens the data?
in this format?
I found savitzky-golay filter, moving average, etc.
But I fail to make it on these kind of data - where x axis is some product = not value.
Please, can you help?
Thanks!!!
d = {'Price': [10, 15, 9, 21,24,23], 'Animal': ['horse', 'lamb', 'gorilla', 'child','panda','duck']}
df = pd.DataFrame(d)
import matplotlib.pyplot as plt
from scipy.optimize import curve_fit
from scipy.interpolate import interp1d
from scipy.signal import savgol_filter
import numpy as np
x = np.arange(1,len(df)+1)
y = df['Price']
xx = np.linspace(x.min(),x.max(), 1001)
# interpolate + smooth
itp = interp1d(x,y, kind='quadratic') #kind = 'linear', 'nearest' (dobre vysledky), slinear (taky ok), cubic (nebrat), quadratic - nebrat
window_size, poly_order = 1001, 1
yy_sg = savgol_filter(itp(xx), window_size, poly_order)
# or fit to a global function
# to stejne jako scipy.optimize.curve.fit
def func(x, A, B, x0, sigma):
return A+B*np.tanh((x-x0)/sigma)
fit, _ = curve_fit(func, x, y)
yy_fit = func(xx, *fit)
fig, ax = plt.subplots(figsize=(7, 4))
ax.plot(x, y, 'r.', label= 'Unsmoothed curve')
ax.plot(xx, yy_fit, 'b--', label=r"$f(x) = A + B \tanh\left(\frac{x-x_0} {\sigma}\right)$")
ax.plot(xx, yy_sg, 'k', label= "Smoothed curve")
plt.legend(loc='best')
I am getting : AttributeError: 'range' object has no attribute 'min'
Savitzky golay is producing very strange values.
With window lenght 1000
When I set window to len(df) +1 (in order it to be odd) then I get these data:
You're getting that error because of the following line:
x = range(1,len(df)).
As the error tells you, a range object has no attribute min.
However, numpy.array()s do, so if you change that line to
x = np.arange(1, len(df)) then this error (at least) will disappear.
EDIT:
In order for the function to do what you want it to do, you should change it to x = np.arange(1, len(df)+1)

How to use Python to draw a normal probability plot by using certain column data in dataFrame

I have a Data Frame that contains two columns named, "thousands of dollars per year", and "EMPLOY".
I create a new variable in this data frame named "cubic_Root" by computing the data in df['thousands of dollars per year']
df['cubic_Root'] = -1 / df['thousands of dollars per year'] ** (1. / 3)
The data in df['cubic_Root'] like that:
ID cubic_Root
1 -0.629961
2 -0.405480
3 -0.329317
4 -0.480750
5 -0.305711
6 -0.449644
7 -0.449644
8 -0.480750
Now! How can I draw a normal probability plot by using the data in df['cubic_Root'].
You want the "Probability" Plots.
So for a single plot, you'd have something like below.
import scipy.stats
import numpy as np
import matplotlib.pyplot as plt
# 100 values from a normal distribution with a std of 3 and a mean of 0.5
data = 3.0 * np.random.randn(100) + 0.5
counts, start, dx, _ = scipy.stats.cumfreq(data, numbins=20)
x = np.arange(counts.size) * dx + start
plt.plot(x, counts, 'ro')
plt.xlabel('Value')
plt.ylabel('Cumulative Frequency')
plt.show()
If you want to plot a distribution, and you know it, define it as a function, and plot it as so:
import numpy as np
from matplotlib import pyplot as plt
def my_dist(x):
return np.exp(-x ** 2)
x = np.arange(-100, 100)
p = my_dist(x)
plt.plot(x, p)
plt.show()
If you don't have the exact distribution as an analytical function, perhaps you can generate a large sample, take a histogram and somehow smooth the data:
import numpy as np
from scipy.interpolate import UnivariateSpline
from matplotlib import pyplot as plt
N = 1000
n = N/10
s = np.random.normal(size=N) # generate your data sample with N elements
p, x = np.histogram(s, bins=n) # bin it into n = N/10 bins
x = x[:-1] + (x[1] - x[0])/2 # convert bin edges to centers
f = UnivariateSpline(x, p, s=n)
plt.plot(x, f(x))
plt.show()
You can increase or decrease s (smoothing factor) within the UnivariateSpline function call to increase or decrease smoothing. For example, using the two you get:
Probability density Function (PDF) of inter-arrival time of events.
import numpy as np
import scipy.stats
# generate data samples
data = scipy.stats.expon.rvs(loc=0, scale=1, size=1000, random_state=123)
A kernel density estimation can then be obtained by simply calling
scipy.stats.gaussian_kde(data,bw_method=bw)
where bw is an (optional) parameter for the estimation procedure. For this data set, and considering three values for bw the fit is as shown below
# test values for the bw_method option ('None' is the default value)
bw_values = [None, 0.1, 0.01]
# generate a list of kde estimators for each bw
kde = [scipy.stats.gaussian_kde(data,bw_method=bw) for bw in bw_values]
# plot (normalized) histogram of the data
import matplotlib.pyplot as plt
plt.hist(data, 50, normed=1, facecolor='green', alpha=0.5);
# plot density estimates
t_range = np.linspace(-2,8,200)
for i, bw in enumerate(bw_values):
plt.plot(t_range,kde[i](t_range),lw=2, label='bw = '+str(bw))
plt.xlim(-1,6)
plt.legend(loc='best')
Reference:
Python: Matplotlib - probability plot for several data set
how to plot Probability density Function (PDF) of inter-arrival time of events?

Plot 95% confidence interval errorbar python pandas dataframes

I want to show 95% confidence interval with Python pandas, matpolib...
But I stucked, because for usual .std() I would do smth like this:
import pandas as pd
import numpy as np
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import math
data = pd.read_table('output.txt',sep=r'\,', engine='python')
Ox = data.groupby(['Ox'])['Ox'].mean()
Oy = data.groupby(['Ox'])['Oy'].mean()
std = data.groupby(['Ox'])['Oy'].std()
plt.plot(Ox, Oy , label = 'STA = '+ str(x))
plt.errorbar(Ox, Oy, std, label = 'errorbar', linewidth=2)
plt.legend(loc='best', prop={'size':9.2})
plt.savefig('plot.pdf')
plt.close()
But I haven't found something in pandas methods which can help me. Does anybody know?
Using 2 * std to estimate the 95 % interval
In a normal distribution, the interval [μ - 2σ, μ + 2σ] covers 95.5 %, so
you can use 2 * std to estimate the 95 % interval:
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
df = pd.DataFrame()
df['category'] = np.random.choice(np.arange(10), 1000, replace=True)
df['number'] = np.random.normal(df['category'], 1)
mean = df.groupby('category')['number'].mean()
std = df.groupby('category')['number'].std()
plt.errorbar(mean.index, mean, xerr=0.5, yerr=2*std, linestyle='')
plt.show()
Result:
Using percentiles
If your distribution is skewed, it is better to use asymmetrical errorbars and get your 95% interval from the percentiles.
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from scipy.stats import skewnorm
df = pd.DataFrame()
df['category'] = np.random.choice(np.arange(10), 1000, replace=True)
df['number'] = skewnorm.rvs(5, df['category'], 1)
mean = df.groupby('category')['number'].mean()
p025 = df.groupby('category')['number'].quantile(0.025)
p975 = df.groupby('category')['number'].quantile(0.975)
plt.errorbar(
mean.index,
mean,
xerr=0.5,
yerr=[mean - p025, p975 - mean],
linestyle='',
)
plt.show()
Result:
For a normal distribution ~95% of the values lie within a window of 4 standard deviations around the mean, or in other words, 95% of the values are within plus/minus 2 standard deviations from the mean. See, e.g. 68–95–99.7-rule.
plt.errorbar's yerr argument specifies the length of the single sided errorbar. Thus taking
plt.errorbar(x,y,yerr=2*std)
where std is the standard deviation shows the errorbars of the 95% confidence interval.
To obtain the 95% confidence interval you'll need to define a function.
Tweaking: Compute a confidence interval from sample data
def mean_confidence_interval(data, confidence=0.95):
a = 1.0 * np.array(data)
n = len(a)
m, se = np.mean(a), scipy.stats.sem(a)
h = se * scipy.stats.t.ppf((1 + confidence) / 2., n-1)
return m
def bound_confidence_interval(data, confidence=0.95):
a = 1.0 * np.array(data)
n = len(a)
m, se = np.mean(a), scipy.stats.sem(a)
h = se * scipy.stats.t.ppf((1 + confidence) / 2., n-1)
return h
Then define:
mean = df.groupby(by='yourquery').agg(mean_confidence_interval)
bound = df.groupby(by='yourquery').agg(bound_confidence_interval)
Finally plot with a library of your choice: e.g. plotly
import plotly.graph_objects as go
fig = go.Figure(data=go.Scatter(
x=mean[yourquery],
y=mean[yourquery2],
error_y=dict(
type='data', # value of error bar given in data coordinates
array=bound[yourquery2],
visible=True)
))
fig.show()

Categories

Resources