classification: Multiclass classification

In AlineTalhouk/splendid: SuPervised Learning ENsemble for Diagnostic IDentification

Multiclass classification

Description

Run a multiclass classification algorithm on a given dataset and reference class.

Usage

classification(
  data,
  class,
  algorithms,
  rfe = FALSE,
  ova = FALSE,
  standardize = FALSE,
  sampling = c("none", "up", "down", "smote"),
  seed_samp = NULL,
  sizes = NULL,
  trees = 100,
  tune = FALSE,
  seed_alg = NULL,
  convert = FALSE
)

Arguments

data	data frame with rows as samples, columns as features
class	true/reference class vector used for supervised learning
algorithms	character string of algorithm to use for supervised learning. See Algorithms section for possible options.
rfe	logical; if TRUE, run Recursive Feature Elimination as a feature selection method for "lda", "rf", and "svm" algorithms.
ova	logical; if TRUE, use the One-Vs-All approach for the knn algorithm.
standardize	logical; if TRUE, the training sets are standardized on features to have mean zero and unit variance. The test sets are standardized using the vectors of centers and standard deviations used in corresponding training sets.
sampling	the default is "none", in which no subsampling is performed. Other options include "up" (Up-sampling the minority class), "down" (Down-sampling the majority class), and "smote" (synthetic points for the minority class and down-sampling the majority class). Subsampling is only applicable to the training set.
seed_samp	random seed used for reproducibility in subsampling training sets for model generation
sizes	the range of sizes of features to test RFE algorithm
trees	number of trees to use in "rf" or boosting iterations (trees) in "adaboost"
tune	logical; if TRUE, algorithms with hyperparameters are tuned
seed_alg	random seed used for reproducibility when running algorithms with an intrinsic random element (random forests)
convert	logical; if TRUE, converts all categorical variables in data to dummy variables. Certain algorithms only work with such limitations (e.g. LDA).

Details

Some of the classification algorithms implemented use pre-defined values that specify settings and options while others need to tune hyperparameters. "multinom" and "nnet" use a maximum number of weights of 2000, in case data is high dimensional and classification is time-consuming. "nnet" also tunes the number of nodes (1-5) in the hidden layer. "pam" considers 100 thresholds when training, and uses a uniform prior. "adaboost" calls maboost::maboost() instead of adabag::boosting() for faster performance. As a result, we use the "entrop" option, which uses the KL-divergence method and mimics adaboost. However, "adaboost_m1" calls adabag::boosting() which supports hyperparameter tuning.

When alg = "knn", the return value is NULL because class::knn() does not output an intermediate model object. The modelling and prediction is performed in one step. However, the class attribute "knn" is still assigned to the result in order to call the respective prediction() method. An additional class "ova" is added if ova = TRUE.

Value

The model object from running the classification algorithm

Algorithms

The classification algorithms currently supported are:

• Prediction Analysis for Microarrays ("pam")

• Support Vector Machines ("svm")

• Random Forests ("rf")

• Linear Discriminant Analysis ("lda")

• Shrinkage Linear Discriminant Analysis ("slda")

• Shrinkage Diagonal Discriminant Analysis ("sdda")

• Multinomial Logistic Regression using

  • Generalized Linear Model with no penalization ("mlr_glm")

  • GLM with LASSO penalty ("mlr_lasso")

  • GLM with ridge penalty ("mlr_ridge")

  • GLM with elastic net penalty ("mlr_enet")

  • Neural Networks ("mlr_nnet")

• Neural Networks ("nnet")

• Naive Bayes ("nbayes")

• Adaptive Boosting ("adaboost")

• AdaBoost.M1 ("adaboost_m1")

• Extreme Gradient Boosting ("xgboost")

• K-Nearest Neighbours ("knn")

Author(s)

Derek Chiu

Examples

data(hgsc)
class <- attr(hgsc, "class.true")
classification(hgsc, class, "xgboost")

AlineTalhouk/splendid documentation built on Feb. 23, 2024, 9:37 p.m.