train_xy.gsom: Train a supervised Growing Self-Organizing Map

In GrowingSOM: Growing Self-Organizing Maps

Description

Computes a supervised growing self-organizing map for mapping high-dimendional data to 2D.

Usage

  train_xy.gsom(data, y, spreadFactor=0.8, keepdata=FALSE, 
      iterations=50, alpha=0.9, beta = 0.5, gridsize = FALSE, 
      nhood= "rect", initrad = NULL, ...)

Arguments

data	a matrix or data.frame, with each row representing an observation and each column a dimension.
y	a matrix or data.frame, with each row representing an observation and each column a y-dimension.
spreadFactor	the spread factor determines the rate with which new units are added to the map. Values close to 0 lead to few growth and therefore less nodes thatn values close to 1. The default value is 0.9.
keepdata	if set to TRUE, a copy of the traindata will be stored in the gsom object.
iterations	number of times that the dataframe will be presented to the network. (Growing and Smoothing Phase combined)
alpha	discount factor for the learning rate during the growing phase of the training. Values should be between 0 and 1.
beta	propagation rate. Determines the rate at which the error of a node, that cannot grow any nodes, is passed on to its neighbours. Suggested values may range between 0 and 1.
gridsize	default value is FALSE. If a nummeric value is entered, the grid-size of the network will be preditermined as a square with length of gridsize. No growth of the network will take place in this case and the values for alpha and beta will be ignored.
nhood	define how the grid will be built, and how the neighbourhood will look like consequently. Allowed values are "rect" (rectangular) and "hex" (hexagonal).
initrad	if the gridsize is predetermined, the initial radius can be chosen here. If left blank, the square root of gridsize will be taken. A larger initrad can increase the quality of the clustering. However, the script can get very slow when a too large value is chosen. (number of necessary computations rises exponentially)
...	not used.

Details

Euclidean distance is used in order to calculate the distances.

Value

an S3 object of class "gsom" with components:

nodes$position	the location of the nodes on the map.
nodes$codes	codes that were established during the training for each node and dimension of the data.
nodes$predict	codes for the y-variables that were established during the training for each node and dimension of y.
nodes$distance	average distance of observations from their best matching units, per unit.
nodes$freq	how many times each node was the best matching unit during the last iteration.
training	reports on the progress of the training after each iteration over the data. "training_stage" indicates whether the algotithm was in the growing phase (1) or the smoothing phase (2), "meandist" records the average distance to the best matching unit, "num_of_nodes" stores the size of the map, and "nodegrowth" keeps track of the nodes grown during each iteration.
GT	the accumulated error per node that was required to grow a new unit during the training phase of the network.
norm_param	parameters which were used for each dimension to normalize the data. Needed to map/predict future data.
norm_param_y	parameters which were used for each y-dimension to normalize the data. Needed to predict future data.
data	data that was used to train the model, not normalized.

Note

In contrast to the paper the algorighm is based on the following adjustments have been made:

1) Changing the Formula to calculate the growth-rate from the SpreadingFactor, in order to take into consideration the number of observations

2) The learning rate is reduced according to the formula used for regualr Kohonen networks during phase tow, in order to prevent too fast depreciation.

The decrease of neighbourhood size only takes place in phase two and happens on a linear basis from the specified start value to 0.

Author(s)

Alex Hunziker

References

Damminda Alahakoon, Saman K. Halgamuge (2000): Dynamic Self-Organizing Maps with Controlled Growth for Knowledge Discovery. IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 11.

See Also

train.gsom, predict.gsom, plot.gsom

Examples

  # load data
  data("auto_mpg")
  s = sample(1:392, 300)
  train_set = auto_mpg[s,1:8]

  # Train Gsom Model
  gsom_object <- train_xy.gsom(train_set[,2:8], train_set[,1], spreadFactor = 0.9)

  # Fixed Grid size
  gsom_object <- train_xy.gsom(train_set[,2:8], train_set[,1], gridsize = 6)

  # Train Gsom Model (hexagonal grid)
  gsom_object <- train_xy.gsom(train_set[,2:8], train_set[,1], spreadFactor = 0.9, nhood="hex")

  # Plot Predicted values for each node
  plot(gsom_object)
  par(mfrow=c(2,2))
  plot(gsom_object, type = "property")
  par(mfrow=c(1,1))
  plot(gsom_object, type = "predict")

GrowingSOM documentation built on May 30, 2017, 6:24 a.m.