Evaluating Self-Organizing Map Quality Measures as Convergence Criteria
The self-organizing map (SOM) is a type of artificial neural network that has applications in a variety of fields and disciplines. The SOM algorithm uses unsupervised learning to produce a low-dimensional representation of high-dimensional data by "fitting'' a grid of nodes to the data over a fixed number of iterations. The low-dimensionality of the resulting map allows for a graphical presentation of the data which can be easily interpreted by humans. To ensure that these models are indeed representative of the underlying data, it is essential to evaluate the quality of the maps. Various measures have been developed that quantify a maps' preservation of topology and neighborhoods. Little work, however, has been done comparing these measures to one another. To that end, this research shows that the quality measures used with SOM can be evaluated as convergence criteria. This is achieved by examining the underlying structure of maps that are converged under different measures. Specifically, the clusters that exist in the maps are compared with the clusters that exist in the input data. For this research, popular real world and synthetic data sets are used for training. The quality measures studied are quantization error, topographic error, topographic function, neighborhood preservation, and population-based convergence.^
Gregory T Breard,
"Evaluating Self-Organizing Map Quality Measures as Convergence Criteria"
Dissertations and Master's Theses (Campus Access).