A self-organizing map (SOM) is a type of unsupervised artificial neural network . The method projects sparse, coupled, nonlinear, and multidimensional data into a lower dimensional space using vector quantization. Following the initial distribution of random seed vectors (weights) and numerous iterations, the competitive learning process results in a network of information in which topological relationships within the training set are maintained by a neighborhood function. Unlike other types of artificial neural networks, the SOM does not need target output to be specified, and its neighborhood function can be used to impute values based on the organized data vector relations. It is this imputation process that facilitates hindcasting and forecasting in this study.
Was this article helpful?