Selecting the variables that train a self-organizing map (SOM) which best separates predefined clusters

Abstract

The paper presents how to find the variables that best illustrate a problem of interest when visualizing with the self-organizing map (SOM). The user defines what is interesting by labeling data points, e.g. with alphabets. These labels assign the data points into clusters. An optimization algorithm looks for the set of variables that best separates the clusters. These variables reflect the knowledge the user applied when labeling the data points. The paper measures the separability, not in the variable space, but on a SOM trained into this space. The found variables contain interesting information, and are well suited for the SOM. The trained SOM can comprehensively visualize the problem of interest, which supports discussion and learning from data. The approach is illustrated using the case of the Hitura mine; and compared with a standard statistical visualization algorithm, the Fisher discriminant analysis.