One-Layer vs. Two-Layer SOM in the Context of Outlier Identification: A Simulation Study

In this work, we applied a stochastic simulation methodology to quantify the power of detection outlying mixture components model, when applying reduced-dimension clustering technique such as Self-Organizing Maps (SOMs). The essential feature SOMs, besides dimensional reduction into discrete map, is conservation topology. two forms learning are applied: competitive, by sequential allocation sample observations winning node in and cooperative, update weights its neighbors. By means cooperative learning, topology from original data space reduced (typically 2D) map achieved. Here, compared performance one- two-layer SOMs outlier representation task. same stratified sampling was for both one-layer SOMs; although, stratification would only be relevant setting—to estimate component power. Two distance measures between points were defined results experiment showed that setting more efficient while maintaining basic properties SOM, which included adequately representing distances remaining ones.