theta parameter optimisation

Classification parameter optimisation for the KNN implementation of Wu and Dietterich's transfer learning schema

Usage

knntlOptimisation(
  primary,
  auxiliary,
  fcol = "markers",
  k,
  times = 50,
  test.size = 0.2,
  xval = 5,
  by = 0.5,
  length.out,
  th,
  xfolds,
  BPPARAM = BiocParallel::bpparam(),
  method = "Breckels",
  log = FALSE,
  seed
)

Arguments

primary: An instance of class "MSnSet".
auxiliary: An instance of class "MSnSet".
fcol: The feature meta-data containing marker definitions. Default is markers.
k: Numeric vector of length 2, containing the best k parameters to use for the primary (k[1]) and auxiliary (k[2]) datasets. See knnOptimisation for generating best k.
times: The number of times cross-validation is performed. Default is 50.
test.size: The size of test (validation) data. Default is 0.2 (20 percent).
xval: The number of rounds of cross-validation to perform.
by: The increment for theta, must be one of c(1, 0.5, 0.25, 0.2, 0.15, 0.1, 0.05)
length.out: Alternative to using by parameter. Specifies the desired length of the sequence of theta to test.
th: A matrix of theta values to test for each class as generated from the function thetas, the number of columns should be equal to the number of classes contained in fcol. Note: columns will be ordered according to getMarkerClasses(primary, fcol). This argument is only valid if the default method 'Breckels' is used.
xfolds: Option to pass specific folds for the cross validation.
BPPARAM: Required for parallelisation. If not specified selects a default BiocParallelParam, from global options or, if that fails, the most recently registered() back-end.
method: The k-NN transfer learning method to use. The default is 'Breckels' as described in the Breckels et al (2016). If 'Wu' is specificed then the original method implemented Wu and Dietterich (2004) is implemented.
log: A logical defining whether logging should be enabled. Default is FALSE. Note that logging produes considerably bigger objects.
seed: The optional random number generator seed.

Value

A list of containing the theta combinations tested, associated macro F1 score and accuracy for each combination over each round (specified by times).

Details

knntlOptimisation implements a variation of Wu and Dietterich's transfer learning schema: P. Wu and T. G. Dietterich. Improving SVM accuracy by training on auxiliary data sources. In Proceedings of the Twenty-First International Conference on Machine Learning, pages 871 - 878. Morgan Kaufmann, 2004. A grid search for the best theta is performed.

References

Breckels LM, Holden S, Wonjar D, Mulvey CM, Christoforou A, Groen AJ, Kohlbacher O, Lilley KS, Gatto L. Learning from heterogeneous data sources: an application in spatial proteomics. bioRxiv. doi: http://dx.doi.org/10.1101/022152

Wu P, Dietterich TG. Improving SVM Accuracy by Training on Auxiliary Data Sources. Proceedings of the 21st International Conference on Machine Learning (ICML); 2004.

Author