Implementation of the KSU compression algorithm https://www.cs.bgu.ac.il/~karyeh/compression-arxiv.pdf

## Project description

## KSU Compression Algorithm Implementation ##

Algortihm 1 from [Nearest-Neighbor Sample Compression: Efficiency, Consistency, Infinite Dimensions](https://arxiv.org/abs/1705.08184)

Installation
------------
* With pip: `pip install ksu`
* From source:
* `git clone --depth=1 https://github.com/nimroha/ksu_classifier.git`
* `cd ksu_classifier`
* `python setup.py install`

Usage
-----

##### Command Line #####

This package provides two command line tools: `e-net` and `ksu`:

* `e-net` constructs an [epsilon net](https://en.wikipedia.org/wiki/Delone_set) for a given epsilon
* `ksu` runs the full algorithm

Both provide the -h flag to specify the arguments, and both can save the result to the disk in [numpy's .npz](https://docs.scipy.org/doc/numpy/reference/generated/numpy.savez.html) format

<br>

##### Code #####

This package provides a class `KSU(Xs, Ys, metric, [gram, prune, logLevel, n_jobs])`

`Xs` and `Ys` are the data points and their respective labels as [numpy arrays](https://docs.scipy.org/doc/numpy/reference/generated/numpy.array.html)

`metric` is either a callable to compute the metric or a string that names one of our provided metrics (print `KSU.METRICS.keys()` for the full list)

`gram` _(optional, default=None)_ a precomputed [gramian matrix](http://mathworld.wolfram.com/GramMatrix.html), will be calculated if not provided.

`prune` _(optional, default=False)_ a boolean indicating whether to prune the compressed set or not (Algorithm 2 from [Near-optimal sample compression for nearest neighbors](https://arxiv.org/abs/1404.3368))

`logLevel` _(optional, default='CRITICAL')_ a string indicating the logging level (set to 'INFO' or 'DEBUG' to get more information)

`n_jobs` _(optional, default=1)_ an integer defining how many cpus to use (scipy logic), pass -1 to use all. For n_jobs below -1, (n_cpus + 1 + n_jobs) are used. Thus for n_jobs = -2, all CPUs but one are used.

<br>

`KSU` provides a method `compressData([delta, minCompress, maxCompress, greedy, stride, logLevel, numProcs])`

Which selects the subset with the lowest estimated error with confidence `1 - delta`. Can take arguments:

`delta` _(optional, default=0.1)_ confidence for error upper bound

`minCompress` _(optional, default=0.05)_ minimal compression ratio

`maxCompress` _(optional, default=0.1)_ maximum compression ratio

`greedy` _(optional, default=True)_ whether to use greedy or hierarichal strategy for net construction

`stride` _(optional, default=200)_ how many gammas to skip between each iteration (since similar gammas will produce similar nets)

`logLevel` _(optional, default='CRITICAL')_ a string indicating the logging level (set to 'INFO' or 'DEBUG' to get more information)

`numProcs` _(optional, default=1)_ number of processes to use

<br>

You can then run `getClassifier()` which returns a 1-NN Classifer (based on [sklearn's K-NN](http://scikit-learn.org/stable/modules/generated/sklearn.neighbors.KNeighborsClassifier.html)) fitted to the compressed data.

Or, run `getCompressedSet()` to get the compressed data as a tuple of numpy arrays `(compressedXs, compressedYs)`.

<br>

See `scripts/` for example usage

##### Built-in metrics #####

['chebyshev', 'yule', 'sokalmichener', 'canberra', 'EarthMover', 'rogerstanimoto', 'matching', 'dice', 'EditDistance', 'braycurtis', 'russellrao', 'cosine', 'cityblock', 'l1', 'manhattan', 'sqeuclidean', 'jaccard', 'seuclidean', 'sokalsneath', 'kulsinski', 'minkowski', 'mahalanobis', 'euclidean', 'l2', 'hamming', 'correlation', 'wminkowski']

## Project details

This version 0.3.0 0.2.2 0.2.1 0.2.0 0.1.2 0.1.1 0.1.0