next up previous contents
Next: Map visualization Up: A Data Mining Tool Previous: Grasping the meaning:

Map training

ENTIRE offers a set of basic training options. The map structure is always a 2-dimensional regular lattice, based either on a hexagonal or a rectangular grid. Of course by setting either of the grid dimensions to one, a 1-dimensional map can be obtained.

For initialization ENTIRE offers two basic algorithms: random initialization with small random values and linear initialization along the subspace spanned by the two principal eigenvectors of the data set (see section 2.2). In the training phase, the neighborhood function can be either bubble or gaussian and the learning coefficient can be selected to decrease either linearly or proportional to inverse of time. Optionally only a selected set of components can be used to organize the map, as described in section 3.3.2.

Since the training algorithm is not deterministic, but depends on the initial values of the model vectors and on the order in which the input vectors are presented to the map, it is useful to train several maps and select the best one for further use. For comparing different maps the user can calculate the quantization error, or some other quality measure, of the map (see section 3.2). In ENTIRE several such quality measures have been implemented: the average quantization error, the topographic error and topological quantization error.

Juha Vesanto
Tue May 27 12:40:37 EET DST 1997