shuffled-stats 1.0.6

Installation

$ pip install shuffled_stats

Examples (without noise)

Let’s start with some simple examples. We construct some random 2-dimensional input data, and the corresponding labels. We then apply shuffled regression using the shuffled_stats.linregress function.

import numpy as np, shuffled_stats

np.random.seed(1)

x = np.random.normal(1, 1, (100,2)) #input features
y = 3*x[:,0] - 7*x[:,1] #labels

np.random.shuffle(y) #in-place shuffling of the labels

shuffled_stats.linregress(x,y) #performs shuffled linear regression
>>> array([3., -7.])

The original weights, [3, -7], are recovered exactly.

We can do another example with defined data points:

Standard linear regression would clearly reveal that 1*x1 + 1*x2 = y. Let’s see what shuffled linear regression reveals with a permuted version of the labels:

x = np.array([[1,2],[2,5],[-1, -2],[5,5],[2,10]])
y = np.array([-3, 10, 7, 3, 12])

shuffled_stats.linregress(x,y) #performs shuffled linear regression
>>> array([1., 1.])

Again, the weights are recovered exactly.

Examples (with noise)

np.random.seed(1) #for reproducibility
x = np.random.normal(1, 1, (100,3)) #input features
x[:,0] = 1 #making a bias/intercept column
y = 4 + 2*x[:,1] - 3*x[:,2] #labels

y = y + np.random.normal(0, .3, (100)) #adding Gaussian noise

w = shuffled_stats.linregress(x,y)
np.round(w,2)
>>> array([3.80, 2.09, -2.91])

We see that the recovered weights approximate the original weights (4, 2, -3), including the bias term.

The library includes a function, shuffled_stats.generate_dataset to quickly generating datasets for testing. Here’s an example:

np.random.seed(1) #for reproducibility

x, y, w0 = shuffled_stats.generate_dataset(n=100, dim=3, bias=True, noise=0.3, mean=2)
w = shuffled_stats.linregress(x,y)

print(np.round(w0,2))
>>> array([2.07, -1.47, -0.83])
print(np.round(w,2))
>>> array([1.79, 1.55, -0.63])

The weights are approximately recovered. We can quantify the relative error by using shuffled_stats.error_in_weights.

shuffled_stats.error_in_weights(w0,w)
>>> 0.13010948373615697 #13% error

Can we improve performance by running three separate “trials” or “replications” of this experiment, each consisting of 100 unordered labels (within each trial, the ordering of the labels is unknown, but labels within a trial must correspond to data points from that trial)? We can test this easily with our library:

np.random.seed(1) #for reproducibility
x, y, w0, groups = shuffled_stats.generate_dataset(n=300, dim=3, weights=[2.07, -1.47, -0.83], bias=True, noise=0.3, mean=2, n_groups=3) #fix weights to the same values as before
w = shuffled_stats.linregress(x,y, groups=groups)

print(np.round(w,2))
>>> array([2.09, -1.48, -0.83])
shuffled_stats.error_in_weights(w0,w)
>>> 0.0099665304764283077 #<1% error

The weights are a lot closer this time!

The library includes several different estimators (see paper for details). We can choose different estimators to compare results:

np.random.seed(1) #for reproducibility
x, y, w0 = shuffled_stats.generate_dataset(n=100, dim=3, weights=[1,1,1], noise=0.3, mean=1) #the true weights are [1,1]
w = shuffled_stats.linregress(x,y, estimator='SM')
print(np.round(w,2))
>>> [0.98  0.98  1.03]
w = shuffled_stats.linregress(x,y, estimator='LS')
print(np.round(w,2))
>>> [0.99  0.92  1.09]
w = shuffled_stats.linregress(x,y, estimator='EMD')
print(np.round(w,2))
>>> [0.99  0.93  1.09]

Examples (on datasets)

Finally, we include methods to load datasets from .csv files (shuffled_stats.load_dataset_in_clusters) so that the performance of shuffled regression can be compared to that of, for example, ordinary least-squares, on real-world data. Here’s an example that uses the accidents.csv dataset, from the UCI repository.

from sklearn.linear_model import LinearRegression

np.random.seed(1) #for reproducibility

x, y, groups = shuffled_stats.load_dataset_in_clusters('accidents.csv', normalize=True, n_clusters = 2)

lr = LinearRegression(fit_intercept=False) #fit_intercept is false because x already includes a bias column

print(lr.fit(x,y).coef_)
>>> [1.02859104,  0.03967381]

print(shuffled_stats.linregress(x,y))
>>> [ 1.12348216  0.02539006]

Not bad, if I do say so myself! Feel free to explore shuffled regression and reach out to me if you have any questions!