Using Bayesian Blocks to Partition Self-Organizing Maps

TL;DR

This paper introduces a novel partitioning scheme for self-organizing maps using Bayesian Blocks segmentation, which effectively identifies boundaries between regions without relying on assumptions about cluster structure.

Contribution

The paper presents a new partitioning method for SOMs based on Bayesian Blocks, offering advantages over traditional methods by minimizing a well-defined cost function.

Findings

Effective boundary detection in SOMs

Independent of cluster assumptions

Sample code provided

Abstract

Self organizing maps (SOMs) are widely-used for unsupervised classification. For this application, they must be combined with some partitioning scheme that can identify boundaries between distinct regions in the maps they produce. We discuss a novel partitioning scheme for SOMs based on the Bayesian Blocks segmentation algorithm of Scargle [1998]. This algorithm minimizes a cost function to identify contiguous regions over which the values of the attributes can be represented as approximately constant. Because this cost function is well-defined and largely independent of assumptions regarding the number and structure of clusters in the original sample space, this partitioning scheme offers significant advantages over many conventional methods. Sample code is available.