Data Driven Resource Discovery Using Self-Organising Maps - An Introduction

The self-organising map (SOM) is an exploratory data mining technique that is both non-traditional and underutilised. Methodologies based on SOM tend to be data-driven and unsupervised, which makes them ideal to assist in the integrated analysis and interpretation of complex and disparate æmineral explorationÆ data sets. While traditional statistical multivariate approaches have difficulty with relationships that are non-linear and data distributions that are not normal, SOM-based data mining procedures are useful in these circumstances. In a SOM analysis each sample is treated as a vector in a data space defined by the variables; and measures of vector similarity, such as the dot-product or Euclidean distance, are used to order or segment a data set into naturally occurring populations. These groupings are positioned as nodes, or groups of nodes, on a 2D rectilinear representation of the multi-dimensional ædata spaceÆ, which is the æself organized mapÆ. While it is common not to include a sampleÆs locational information in the actual SOM analysis, such information can be used to display the spatial location of samples coded by their SOM node or cluster. If one finds there are coherent spatial patterns belonging to samples from particular nodes, or groups of nodes, then there is strong evidence that the technique is determining patterns and relationships within the data that have natural significance. Variations of the SOM technique may be used to perform broad categories of operations, including: 1. outlier identification û for samples belonging to various nodes or grouping of nodes; 2. function fitting, prediction, estimation or imputation û for calibrating relationships between variables and determining replacement values for missing, null or censored values; 3. clustering, pattern recognition or noise reduction û for determining the natural patterns within a data set; and 4. classification û for classifying samples into the clusters determined in three. An EXTENDED ABSTRACT is available for download. A full-length paper was not prepared for this presentation.