Self-Organizing Maps (SOM), also known as Kohonen maps, are a type of unsupervised artificial neural network used for dimensionality reduction, data visualization, and clustering. SOMs are particularly useful for visualizing high-dimensional data in lower-dimensional spaces while preserving the topological relationships between data points.
SOMs consist of a grid of nodes or neurons, where each node represents a weight vector in the input space. During training, the SOM "learns" to arrange these nodes in a way that reflects the similarities and relationships between data points.
The training process steps…
-
Initialization: The weight vectors of the nodes are initialized randomly or using a method like Pricipal Component Analysis (PCA) to capture the principal components of the data.
-
Training (Competitive Learning): In this step, for each input data point, the node with the closest weight vector (i.e., the "best matching unit" or BMU) is identified. The BMU and its neighboring nodes are updated to become more similar to the input data point. The extent of influence on neighboring nodes decreases with distance from the BMU. This process is repeated for all data points over multiple iterations, allowing the SOM to gradually adapt to the data distribution.
The result is a topological representation of the input space, where similar data points are mapped close to each other on the SOM grid. Neighboring nodes in the SOM grid tend to represent similar features or clusters in the data. SOMs can also be used to identify outliers or data points that deviate significantly from the rest. Interpreting the resulting SOM may require domain knowledge to understand the relationships between the nodes and the original data points accurately.