Self organizing feature maps is a neural network method for cluster analysis. A neural network is a set of connected input/output units, where each connection has a weight associated with it. They are popular for clustering because 1) they are inherently parallel and distributed processing architectures 2) they learn by adjusting their inter-connection weights so as to best fit the data. With this, they normalize the patterns and act as feature extractors for the various clusters. 3) They process numerical vectors and require object patterns to be represented by quantitative patterns.
Each cluster is represented as an exemplar which means a prototype and does not have to match a data example. Data points are assigned to cluster that is most similar to an exemplar based on a distance measure. The attributes are then predicted from the attributes of the exemplar.
Self-organizing feature maps represent all points in a high-dimensional source space by a points in 2-D or 3-D space such that distance and closeness are maintained. The method is useful when the problem is inherently non-linear.
SOM can be viewed as constrained versions of k-means clustering where the cluster centers are in low dimensional space.
Clustering is performed with several units competing for the current object. The unit whose weight vector is closest to the current object becomes the winning unit. The weights of this unit and those of its neighbors are adjusted to get closer to the input object. The assumption is there is a topology or ordering in the input that the units will eventually take shape. This is called a feature map. Such processing is applied to web mining but is costly for large databases.
Each cluster is represented as an exemplar which means a prototype and does not have to match a data example. Data points are assigned to cluster that is most similar to an exemplar based on a distance measure. The attributes are then predicted from the attributes of the exemplar.
Self-organizing feature maps represent all points in a high-dimensional source space by a points in 2-D or 3-D space such that distance and closeness are maintained. The method is useful when the problem is inherently non-linear.
SOM can be viewed as constrained versions of k-means clustering where the cluster centers are in low dimensional space.
Clustering is performed with several units competing for the current object. The unit whose weight vector is closest to the current object becomes the winning unit. The weights of this unit and those of its neighbors are adjusted to get closer to the input object. The assumption is there is a topology or ordering in the input that the units will eventually take shape. This is called a feature map. Such processing is applied to web mining but is costly for large databases.
No comments:
Post a Comment