kiburmsong / scikit-sos

A Python implementation of the Stochastic Outlier Selection algorithm

scikit-sos

scikit-sos is a Python module for Stochastic Outlier Selection (SOS). It is compatible with scikit-learn. SOS is an unsupervised outlier selection algorithm. It uses the concept of affinity to compute an outlier probability for each data point.

For more information about SOS, see the technical report: J.H.M. Janssens, F. Huszar, E.O. Postma, and H.J. van den Herik. Stochastic Outlier Selection. Technical Report TiCC TR 2012-001, Tilburg University, Tilburg, the Netherlands, 2012.

Install

pip install scikit-sos

Usage

>>> import pandas as pd
>>> from sksos import SOS
>>> iris = pd.read_csv("http://bit.ly/iris-csv")
>>> X = iris.drop("Name", axis=1).values
>>> detector = SOS() 
>>> iris["score"] = detector.predict(X)
>>> iris.sort_values("score", ascending=False).head(10)
     SepalLength  SepalWidth  PetalLength  PetalWidth             Name     score
41           4.5         2.3          1.3         0.3      Iris-setosa  0.981898
106          4.9         2.5          4.5         1.7   Iris-virginica  0.964381
22           4.6         3.6          1.0         0.2      Iris-setosa  0.957945
134          6.1         2.6          5.6         1.4   Iris-virginica  0.897970
24           4.8         3.4          1.9         0.2      Iris-setosa  0.871733
114          5.8         2.8          5.1         2.4   Iris-virginica  0.831610
62           6.0         2.2          4.0         1.0  Iris-versicolor  0.821141
108          6.7         2.5          5.8         1.8   Iris-virginica  0.819842
44           5.1         3.8          1.9         0.4      Iris-setosa  0.773301
100          6.3         3.3          6.0         2.5   Iris-virginica  0.765657

Selecting outliers from the command line

This module also includes a command-line tool called sos. To illustrate, we apply SOS with a perplexity of 10 to the Iris dataset:

$ curl -sL http://bit.ly/iris-csv |
> tail -n +2 | cut -d, -f1-4 |
> sos -p 10 |
> sort -nr | head
0.98189840
0.96438132
0.95794492
0.89797043
0.87173299
0.83161045
0.82114072
0.81984209
0.77330148
0.76565738

Adding a threshold causes SOS to output 0s and 1s instead of outlier probabilities. If we set the threshold to 0.8 then we see that out of the 150 data points, 8 are selected as outliers:

$ curl -sL http://bit.ly/iris-csv |
> tail -n +2 | cut -d, -f1-4 |
> sos -p 10 -t 0.8 |
> paste -sd+ | bc
8

License

All software in this repository is distributed under the terms of the BSD Simplified License. The full license is in the LICENSE file.