In jlsuarezdiaz/rDML: Distance Metric Learning Algorithms for R

First, we load the needed libraries.

library(datasets)
library(rDML)

We will use the iris dataset.

# Loading dataset
data(iris)
X = iris[1:4]
y = as.array(as.numeric(iris[5][,1]))

A tune function is available for parameter estimation in DML algorithms. We can choose the algorithm to tune giving it string abbreviation. We can set some fixed parameters in the argument dml_params. The parameters to estimate are added in tune_args, as a named list. Each entry has as name a parameter of the DML, and its value is a list with all the values to estimate. The metrics used to measure the algorithm performances can be either different k-NN classifications, or any of the algorithm numeric metadata. The estimation will be done using cross validation.

  result_list = dml_tune$tune(dml = "NCA",X = X,y = y, dml_params = list(descent_method = "SGD"), 
                              tune_args = list(eta0 = c(0.3,0.1,0.01), num_dims = as.integer(c(2,4))), 
                              metrics = list("final_expectance", 3,5),n_folds = 5,n_reps = 2,
                              verbose = TRUE, seed = 28)

We obtain many information after tuning the algorithm:

  results = result_list[[1]]
  best = result_list[[2]]
  nca_best = result_list[[3]]
  detailed = result_list[[4]]

From results we obtain the average cross validation results:

results

In best we can isolate the best result (according to the first metric specified):

best

Also, we can take the algorithm with the best parameters ready to be fitted.

nca_best$fit(X,y)
nca_best$transform()[1:5,]

Finally, we can also look at the detailed results for each fold in the cross validation, with detailed:

  detailed$`{'eta0': 0.3, 'num_dims': 2}`

jlsuarezdiaz/rDML documentation built on May 24, 2019, 12:35 a.m.