scikit-hubness

scikit-hubness provides tools for the analysis and reduction of hubness in high-dimensional data. Hubness is an aspect of the curse of dimensionality and is detrimental to many machine learning and data mining tasks.

The skhubness.analysis and skhubness.reduction packages allow to

• analyze, whether your data sets show hubness
• reduce hubness via a variety of different techniques
• perform downstream analysis (performance assessment) with scikit-learn due to compatible data structures

The skhubness.neighbors package provides approximate nearest neighbor (ANN) search. This is compatible with scikit-learn classes and functions relying on neighbors graphs due to compliance with KNeighborsTransformer APIs and data structures. Using ANN can speed up many scikit-learn classification, clustering, embedding and other methods, including: - KNeighborsClassifier - DBSCAN - TSNE - and many more.

scikit-hubness thus provides - approximate nearest neighbor search - hubness reduction - and combinations,

which allows for fast hubness-reduced neighbor search in large datasets (tested with >1M objects).

Installation

Make sure you have a working Python3 environment (at least 3.8).

Use pip to install the latest stable version of scikit-hubness from PyPI:

bash pip install scikit-hubness

NOTE: v0.30 is currently under development and not yet available on PyPI. Install from sources to obtain the bleeding edge version.

Dependencies are installed automatically, if necessary. scikit-hubness is based on the SciPy-stack, including numpy, scipy and scikit-learn. Approximate nearest neighbor search and approximate hubness reduction additionally require at least one of the following packages: * nmslib for hierachical navigable small-world graphs in skhubness.neighbors.NMSlibTransformer * ngtpy for nearest neighbor graphs (ANNG, ONNG) in skhubness.neighbors.NGTTransformer * puffinn for locality-sensitive hashing in skhubness.neighbors.PuffinnTransformer * annoy for random projection forests in skhubness.neighobrs.AnnoyTransformer * Additional ANN libraries might be added in future releases. Please reach out to us in a Github Issue, if you think a specific library is missing (pull requests welcome).

For more details and alternatives, please see the Installation instructions.

Documentation

Additional documentation is available online: http://scikit-hubness.readthedocs.io/en/latest/index.html

What's new

See the changelog to find what's new in the latest package version.

Quickstart

Users of scikit-hubness may want to

1. analyse, whether their data show hubness
2. reduce hubness
3. perform learning (classification, regression, ...)

The following example shows all these steps for an example dataset from the text domain (dexter). (Please make sure you have installed scikit-hubness).

```python from sklearn.modelselection import crossval_score from sklearn.neighbors import KNeighborsClassifier, KNeighborsTransformer

from skhubness import Hubness from skhubness.data import load_dexter from skhubness.neighbors import NMSlibTransformer from skhubness.reduction import MutualProximity

load the example dataset 'dexter' that is embedded in a

high-dimensional space, and could, thus, be prone to hubness.

X, y = load_dexter() print(f'X.shape = {X.shape}, y.shape = {y.shape}')

assess the actual degree of hubness in dexter

hub = Hubness(k=10, metric='cosine') hub.fit(X) kskew = hub.score() print(f'Skewness = {kskew:.3f}')

additional hubness indices are available, for example:

hub = Hubness(k=10, returnvalue="all", metric='cosine') scores = hub.fit(X).score() print(f'Robin hood index: {scores.get("robinhood"):.3f}') print(f'Antihub occurrence: {scores.get("antihuboccurrence"):.3f}') print(f'Hub occurrence: {scores.get("hub_occurrence"):.3f}')

There is considerable hubness in dexter. Let's see, whether

hubness reduction can improve kNN classification performance.

We first create a kNN graph:

knn = KNeighborsTransformer(n_neighbors=50, metric="cosine")

Alternatively, create an approximate KNeighborsTransformer, e.g.,

knn = NMSlibTransformer(n_neighbors=50, metric="cosine")

kneighborsgraph = knn.fittransform(X, y)

vanilla kNN without hubness reduction

clf = KNeighborsClassifier(nneighbors=5, metric='precomputed') accstandard = crossvalscore(clf, kneighbors_graph, y, cv=5)

kNN with hubness reduction (mutual proximity) reuses the

precomputed graph and works in sklearn workflows:

mp = MutualProximity(method="normal") mpgraph = mp.fittransform(kneighborsgraph) accmp = crossvalscore(clf, mp_graph, y, cv=5)

print(f'Accuracy (vanilla kNN): {accstandard.mean():.3f}') print(f'Accuracy (kNN with hubness reduction): {accmp.mean():.3f}')

Accuracy was considerably improved by mutual proximity.

Did it actually reduce hubness?

mpscores = hub.fit(mpgraph).score() print(f'k-skewness after MP: {mpscores.get("kskewness"):.3f} ' f'(reduction of {scores.get("kskewness") - mpscores.get("kskewness"):.3f})') print(f'Robinhood after MP: {mpscores.get("robinhood"):.3f} ' f'(reduction of {scores.get("robinhood") - mp_scores.get("robinhood"):.3f})') ```

Check the User Guide for additional example usage.

Development

The developers of scikit-hubness welcome all kinds of contributions! Get in touch with us if you have comments, would like to see an additional feature implemented, would like to contribute code or have any other kind of issue. Don't hesitate to file an issue here on GitHub.

For more information about contributing, please have a look at the contributors guidelines.

(c) 2018-2022, Roman Feldbauer
-2018: Austrian Research Institute for Artificial Intelligence (OFAI) and
-2021: University of Vienna, Division of Computational Systems Biology (CUBE)
2021-: Independent researcher
Contact: <sci@feldbauer.org>

Citation

If you use scikit-hubness in your scientific publication, please cite:

@Article{Feldbauer2020,
  author  = {Roman Feldbauer and Thomas Rattei and Arthur Flexer},
  title   = {scikit-hubness: Hubness Reduction and Approximate Neighbor Search},
  journal = {Journal of Open Source Software},
  year    = {2020},
  volume  = {5},
  number  = {45},
  pages   = {1957},
  issn    = {2475-9066},
  doi     = {10.21105/joss.01957},
}

To specifically acknowledge approximate hubness reduction, please cite:

@INPROCEEDINGS{8588814,
author={R. {Feldbauer} and M. {Leodolter} and C. {Plant} and A. {Flexer}},
booktitle={2018 IEEE International Conference on Big Knowledge (ICBK)},
title={Fast Approximate Hubness Reduction for Large High-Dimensional Data},
year={2018},
volume={},
number={},
pages={358-367},
keywords={computational complexity;data analysis;data mining;mobile computing;public domain software;software packages;mobile device;open source software package;high-dimensional data mining;fast approximate hubness reduction;massive mobility data;linear complexity;quadratic algorithmic complexity;dimensionality curse;Complexity theory;Indexes;Estimation;Data mining;Approximation algorithms;Time measurement;curse of dimensionality;high-dimensional data mining;hubness;linear complexity;interpretability;smartphones;transport mode detection},
doi={10.1109/ICBK.2018.00055},
ISSN={},
month={Nov},}

The technical report Fast approximate hubness reduction for large high-dimensional data is available at OFAI.

Additional reading

Local and Global Scaling Reduce Hubs in Space, Journal of Machine Learning Research 2012, Link.

A comprehensive empirical comparison of hubness reduction in high-dimensional spaces, Knowledge and Information Systems 2018, DOI.

License

scikit-hubness is licensed under the terms of the BSD-3-Clause license.

Note: Individual files contain the following tag instead of the full license text.

    SPDX-License-Identifier: BSD-3-Clause

This enables machine processing of license information based on the SPDX License Identifiers that are here available: https://spdx.org/licenses/

Acknowledgements

Parts of scikit-hubness adapt code from scikit-learn. We thank all the authors and contributors of this project for the tremendous work they have done.

PyVmMonitor is being used to support the development of this free open source software package. For more information go to http://www.pyvmmonitor.com