coBC {SSLR}

General Interface for CoBC model

Description

Co-Training by Committee (CoBC) is a semi-supervised learning algorithm with a co-training style. This algorithm trains N classifiers with the learning scheme defined in the learner argument using a reduced set of labeled examples. For each iteration, an unlabeled example is labeled for a classifier if the most confident classifications assigned by the other N-1 classifiers agree on the labeling proposed. The unlabeled examples candidates are selected randomly from a pool of size u. The final prediction is the average of the estimates of the N regressors.

Usage

coBC(learner, N = 3, perc.full = 0.7, u = 100, max.iter = 50)

Arguments

learner

model from parsnip package for training a supervised base classifier using a set of instances. This model need to have probability predictions in classification mode

N

The number of classifiers used as committee members. All these classifiers are trained using the gen.learner function. Default is 3.

perc.full

A number between 0 and 1. If the percentage of new labeled examples reaches this value the self-labeling process is stopped. Default is 0.7.

u

Number of unlabeled instances in the pool. Default is 100.

max.iter

Maximum number of iterations to execute in the self-labeling process. Default is 50.

Details

For regression tasks, labeling data is very expensive computationally. Its so slow. This method trains an ensemble of diverse classifiers. To promote the initial diversity the classifiers are trained from the reduced set of labeled examples by Bagging. The stopping criterion is defined through the fulfillment of one of the following criteria: the algorithm reaches the number of iterations defined in the max.iter parameter or the portion of unlabeled set, defined in the perc.full parameter, is moved to the enlarged labeled set of the classifiers.

Value

(When model fit) A list object of class "coBC" containing:

model

The final N base classifiers trained using the enlarged labeled set.

model.index

List of N vectors of indexes related to the training instances used per each classifier. These indexes are relative to the y argument.

instances.index

The indexes of all training instances used to train the N models. These indexes include the initial labeled instances and the newly labeled instances. These indexes are relative to the y argument.

model.index.map

List of three vectors with the same information in model.index but the indexes are relative to instances.index vector.

classes

The levels of y factor in classification.

pred

The function provided in the pred argument.

pred.pars

The list provided in the pred.pars argument.

References

Avrim Blum and Tom Mitchell.
Combining labeled and unlabeled data with co-training.
In Eleventh Annual Conference on Computational Learning Theory, COLT’ 98, pages 92-100, New York, NY, USA, 1998. ACM. ISBN 1-58113-057-0. doi: 10.1145/279943.279962.

Mohamed Farouk Abdel-Hady, Mohamed Farouk Abdel-Hady and Günther Palm.
Semi-supervised Learning for Regression with Cotraining by Committee
Institute of Neural Information Processing University of Ulm D-89069 Ulm, Germany

Examples

library(tidyverse)
library(tidymodels)
library(caret)
library(SSLR)

data(wine)

set.seed(1)
train.index <- createDataPartition(wine$Wine, p = .7, list = FALSE)
train <- wine[ train.index,]
test  <- wine[-train.index,]

cls <- which(colnames(wine) == "Wine")

#% LABELED
labeled.index <- createDataPartition(wine$Wine, p = .2, list = FALSE)
train[-labeled.index,cls] <- NA

#We need a model with probability predictions from parsnip
#https://tidymodels.github.io/parsnip/articles/articles/Models.html
#It should be with mode = classification

#For example, with Random Forest
rf <-  rand_forest(trees = 100, mode = "classification") %>%
  set_engine("randomForest")


m <- coBC(learner = rf,N = 3,
          perc.full = 0.7,
          u = 100,
          max.iter = 3) %>% fit(Wine ~ ., data = train)

#Accuracy
predict(m,test) %>%
  bind_cols(test) %>%
  metrics(truth = "Wine", estimate = .pred_class)