Overview
Self-organizing maps are a projection technique, which reduces multi-dimensional data into a fixed number of dimensions. This reduced-dimensional view is the map, which is typically two dimensional. This map is constructed of a set of nodes which segment the map, and hence, the data as it is mapped. Where the mapping process has concentrated many records into a single node, it also concentrates the nodes themselves in order to share the records. The resulting map nodes hold and thus represent a set of record that are close neighbors to each other.
Another way of conceptualizing what a self-organizing map does is realizing that the map is fits a n-dimensional object to the extents of the data with preferences towards the areas that are most dense.
Self-organizing maps are most commonly used as a visualization technique for the purposes of clustering or data extents visualization. Self-organizing maps, when degenerate, can duplicate the functionality of the k-means clustering algorithm or the learning vector quantization algorithm. Additionally, the data related to every node can be considered the neighborhood of a k-nearest neighbor classification or regression algorithms.
Typical Application Areas
- Multi-dimensional process control
- process control and state classification
- process tracking and analysis
- Customer data analysis
- segmentation of current customers
- reshaping current groups
- discovering net groups
- common features inside a group
- differentiation between groups
- customer aquisition analysis
- customer retention/churn analysis
- customer profitability analysis
- customer-to-product recommendations
- customer message/marketing optimization
- predicting bad credit and bankruptcies
Feature List:
EWA's Self-Organizing Maps Engine supports maps of n-D (Line, Flat, Cube, and higher-dimensions), of any number of nodes in each dimension, of hex or rectangular layout, and with wrapping applied to any of dimension. The engine supports the building of a single map or a tree SOM, the building of a simple map and then using that to build successively more complex maps. Maps can then be trained with a dot-product or incremental learning functions, and with a variety of vector distance or spreading functions. The resulting map is an excellent visualization tool, and can be used as a nearest-neighbor classification or regression engine, or as a basis for clustering.
- Implemented in 100% Java, but with C-like Performance.
- Unlimited Problem Size (Problem does not have to fit in memory)
- Controlled Multi-Threaded Implementation (Uses only the number of threads specified)
- Uses EWA's Standard Data Manipulation Package
- Configuration options
- Map structure details
- The number of dimension of the map
- The size of, and wrapping related to each map dimension
- Hexagonal or rectangular map pattern
- Tree SOM details
- The number of layers to build
- The magnification applied to sucessive layers
- Vector distance function
- Cartesian
- Longest-link
- City-block
- User definable
- Learning spreading functions
- Linear
- k-Means
- User definable
- Learning rate
- Learning momentum
- Learning model
- Min and max learning iterations
- Resulting model
- The map
- The map's structure
- The map's nodes
- Each nodes associated records
- Visualization
- The map's structure
- U-Matrix (minimum, average, and maximum vector distances)
- Data labels
- Component planes
- The model can be saved in XML, or binary formats
Performance
Comparison tests show the EWA Self-Organizing Maps Engine to have performance similar to that of its C-based competitors, despite being coded in Java.
Self-Organizing Map Engine User's Guide (Request)
Self-Organizing Map Engine JavaDocs (Online/Zipped)
Competitors to EWA Systems' Self-Organizing Maps Engine
To Purchase or For More Information, contact our
Sales Team.