Contact Points:

Andreas Rauber ( ) and Rudolf Mayer ( )
  

Technical Contact Points:

Rudolf Mayer ( ) and Thomas Lidy ( )
  

Type of Software:

PROTOTYPE
  

Descriptive Keywords:

Self-organising map (SOM), Clustering, Visualisation.
  

Potential Use and Applications:

The Java SOMtoolbox can be used to cluster e.g. text, audio, image or other data that can be described in vectorial form. It provides a wealth of tools to inspect the clustering results in an innovative, interactive approach .
  

General Description:

The Java SOMtoolbox provides implementations of Self-organising map models, among them for the classical SOM, the Growing Grid, and the Growing Hierarchical SOM (see http://www.ifs.tuwien.ac.at/mir/howto_matlab_fe.html). The SOM provides a clustering of the high-dimensional input space to an easier understandable two-dimensional output space. Labeling algorithms for the maps are also provided (only applicable if the data has some meaning to extract from the features, e.g. in Text mining). The software can handle any kind of media objects, as long as they are describable in a vectorial format. On top, the toolbox offers a rich desktop client for exploring the trained maps in a novel and interactive way. The interface allows for zooming and panning, and selecting regions, paths or single data objects, to create and export e.g. play lists in case of audio data. A wealth of visualisation techniques is available, e..g the U-Matrix, D-matrix, Component planes, Smoothed Data Histograms (SDH), Class visualisations. Screenshots and more description can be found at http://www.ifs.tuwien.ac.at/mir/#PlaySOM and http://www.ifs.tuwien.ac.at/dm/#GHSOM.
  

Technical Description:

The Self-Organising Map (SOM) is a unsupervised machine learning algorithm that provides a mapping from a high-dimensional input space (e.g. extracted music features) to a low-dimensional, usually two-dimensional, output space. In this mapping, the SOM is topology preserving, meaning that items that are close in the input space will also be closely located in the output space, while objects that are distant in the input space, and therefore dissimilar, will be mapped onto different regions of the SOM. The low-dimensional output space allows for an easier understanding of the underlying cluster structure in the data. Due to this, the SOM is a very interesting method to analyse data, and used in many different applications. With visualisation techniques as the U-Matrix or the SDH, the cluster boundaries can be visualised. The software provides those visualisations, plus the D-Matrix, component planes, visualisations of training quality, class distribution visualisations, and others. More details on the SOM can be found e.g. http://www.ifs.tuwien.ac.at/~andi/ghsom/.
  

Required User Skills:

Handling the command line scripts; if new maps are trained, possible data pre-processing and conversion to the (plain-text) file format needed by the training algorithms.
  

Pre-Requisites for Installation:

25 MB disc space

>= 256 MB RAM for the SOM training

>=512 for the SOM Viewer applications
  

Conditions of Use:

The software may be used for research purposed. Please contact the authors for more information.