OrderStatistics

A repository focused on outlier detection and analysis of heavy-tailed distributions using order statistics. Here you can find:

Simualations of order statistics for outlier detection and the statistic provided in our paper titled On a Notion of Outliers Based on Ratios of Order Statistics.
Single and double bootstrap methods (& more) for tail index estimation of heavy-tailed data.
Special kernel density estimation methods for heavy-tailed data.

Also:

Useful plotting functions for outlier detection.
Easy reporting tools for binary classification such as get_classification_report(), which reports all classification metrics from confusion matrix to AUC and can plot ROC curve.

Installation:

You can install the repository with:

pip install orderstats

Alternatively, you can download a copy of the repository from this page. After downloading, you can do pip install -r requirements.txt to install the requirements.

General explanation about Random Variables:

We use scipy.stats package for generating random variables. A random variable instance can be created just by giving the appropriate parameters. For example for X ~ N(0, 1), we can do:
>>> X = stats.norm(0, 1)
Once an instance is created, we can calculate pdf, or cdf using .pdf(), .cdf() methods, or, we can take a sample using .rvs():
>>> X.pdf(1.96) for pdf of X ~ N(0, 1) at x = 1.96;
>>> X.cdf(1.96) for cdf of X ~ N(0, 1) at x = 1.96;
>>> X.rvs(1000) for an i.i.d sample of 1000 from X ~ N(0, 1).

Some problems:

Not every random variable function in scipy.stats is intuitive.
For instance expon function creates an instance of an exponential distribution. However, if we wish to get an instance of exponential distribution with lambda = 2 (i.e. with pdf f(x) = 2e^(-2x)), then we would need to use expon(0, 1/2). In distributions.py, there are examples given for some popular distributions to clarify any ambiguities.

Examples

Outlier Detection:

Here is an example use of our method for a 1D dataset X:

from orderstats import scoring
from orderstats.distributions import moving_average_unscaled_kappa
from orderstats.plot_utils import plot_anomalies
scores, scores_sorted = scoring.get_anomaly_scores(X, scoring_func=moving_average_unscaled_kappa)
threshold = scoring.get_kappa_threshold(scores_sorted)
predictions = scores > threshold
plot_anomalies(X, predictions=predictions)

Order Statistics Simulation:

In general, we will use OrderSimulation class in distributions for simulations. Any random variable from the scipy.stats package can be given to this class as an argument. For example, if you wish to simulate the sums of first m order statistics from a sample of exponential distribution of size n:

from orderstats import OrderSimulation
simulate_normal_dist = OrderSimulation(stats.expon(0, 1), calculate_S_m)
simulation = simulate_normal_dist(10000, n, m)

For studying change point detection, we provide the MixSimulation class. For getting a mixed sample with corresponding ids:

from orderstats import MixSimulation
simulate_mixture = MixSimulation(dist1=expon(0, 1), dist2=stats.pareto(2.))
mixed_array, idx = simulate_mixture(n1, n2)

Tail Index Estimation:

Most of the methods for tail index estimation mentioned in [1] is implemented in tail_estimation. As an example for the double bootstrap method:

import numpy as np
N = 1000
pareto_sample = np.random.pareto(2, 1000)
sample_to_estimate_index = np.sort(X,)[::-1] # Sort decreasing
double_bootsrap = DoubleBootstrap()
tail_index = double_bootstrap(N, sample_to_estimate_index)

References:

[1] Markovich, N. (2008). Nonparametric analysis of univariate heavy-tailed data: research and practice (Vol. 753). John Wiley & Sons.

Selozhd / orderstats