autocluster

autocluster is an automated machine learning (AutoML) toolkit for performing clustering tasks.

Report and presentation slides can be found here and here.

Prerequisites

• Python 3.5 or above
• Linux OS, or Windows WSL is also possible

How to get started?

1. First, install SMAC:

• sudo apt-get install build-essential swig
• conda install gxx_linux-64 gcc_linux-64 swig
• pip install smac==0.8.0

2. pip install autocluster

How it works?

• autocluster automatically optimizes the configuration of a clustering problem. By configuration, we mean

  • choice of dimension reduction algorithm
  • choice of clustering model
  • setting of dimension reduction algorithm's hyperparameters
  • setting of clustering model's hyperparameters

• autocluster provides 3 different approaches to optimize the configuration (with increasing complexity):

  • random optimization
  • bayesian optimization
  • bayesian optimization + meta-learning (warmstarting)

Algorithms/Models supported

• List of dimension reduction algorithms in sklearn supported by autocluster's optimizer.

• List of clustering models in sklearn supported by autocluster's optimizer.

Examples

Examples are available in these notebooks.

Experimental results

• This dataset comprises of 16 Gaussian clusters in 128-dimensional space with N = 1024 points. The optimal configuration obtained by autocluster (SMAC + Warmstarting) consists of a Truncated SVD dimension reduction model + Birch clustering model.

• This dataset comprises of 15 Gaussian clusters in 2-dimensional space with N = 5000 points. The optimal configuration obtained by autocluster (SMAC + Warmstarting) consists of a TSNE dimension reduction model + Agglomerative clustering model.

Links

• Link to pypi.
• Great writeup by Martin Krasser on Bayesian Optimization

Disclaimer

The project is experimental and still under development.