Skip to content

Latest commit

 

History

History
160 lines (131 loc) · 4.94 KB

Caret.md

File metadata and controls

160 lines (131 loc) · 4.94 KB
Raw

Machine Learning with Caret

Jesse Cambon 22 November, 2019

Demonstrate a machine learning workflow with caret

References

Setup

library(mlbench) # machine learning reference datasets
library(tidyverse)
library(broom)
library(caret)
library(skimr)
library(knitr)
library(kableExtra)

data(BreastCancer) # load

# Set seed for reproducibility
set.seed(45)

Build Model

#skim(BreastCancer)

BC <- BreastCancer %>% as_tibble() %>%
  dplyr::select(-Id) %>%
  # should really use imputation but we'll just drop nas for now
  drop_na() 

# Use k-fold cross-validation
TC <- trainControl(method="cv", number=5)

# Train some models

# Neural Net
nnet <- train(Class ~ . , BC,method='nnet',trControl=TC,verbose=FALSE)

# Gradient Boosted Machine
gbm <- train(Class ~ . , BC,method='gbm',trControl=TC)

# Radial SVM
svmrad <- train(Class ~ . , BC,method='svmRadial',trControl=TC)

# Elastic-net
glmnet <- train(Class ~ . , BC,method='glmnet',trControl=TC,tuneLength=5)

# Logistic regression - did not converge
glm <- train(Class ~ . , BC,method='glm',trControl=TC)
# Look at results of Glmnet model

# Extract coefficients from optimal model
glm_coeff <- coef(glmnet$finalModel,glmnet$finalModel$lambdaOpt) %>% 
  as.matrix() %>% as.data.frame() %>%
  rownames_to_column('Variable') %>%
  as_tibble() %>%
  rename(Coefficient=2) %>%
  arrange(desc(abs(Coefficient)))


# Combine variable importance data with coefficients
varImportance <- varImp(glmnet)$importance %>% 
  rownames_to_column('Variable') %>%
  rename(Importance=2) %>%
  arrange(desc(Importance)) %>%
  full_join(glm_coeff,by='Variable') %>%
  filter(Coefficient != 0) 
resamps <- resamples(list(nnet=nnet,
                          glmnet=glmnet,
                          svmrad=svmrad,
                          gbm=gbm,
                          glm=glm))

# Accuracy comparison
dotplot(resamps, metric = "Accuracy")

# Difference in accuracy
bwplot(diff(resamps))

Glmnet (Elastic Net) Model

kable(varImportance,format='markdown') %>%
  kable_styling(bootstrap_options = c("striped",'border'))
Variable Importance Coefficient
Cl.thickness.L 100.0000000 3.1361533
Bare.nuclei9 80.1349380 2.5131545
Bare.nuclei6 72.1692163 2.2633373
Bare.nuclei10 62.8228881 1.9702221
Cell.shape.L 60.0936317 1.8846284
Marg.adhesion.L 59.9667240 1.8806484
Cell.size.L 54.3790530 1.7054105
Normal.nucleoli10 51.3425770 1.6101819
Normal.nucleoli9 48.7621790 1.5292567
Bl.cromatin5 42.9191401 1.3460100
Marg.adhesion^9 38.8206640 1.2174755
Normal.nucleoli4 38.0200861 1.1923682
Cell.shape.Q 34.0212706 -1.0669592
Cl.thickness^8 27.9829503 -0.8775882
Normal.nucleoli2 27.4707392 -0.8615245
Epith.c.size^4 27.1064505 0.8500998
Bare.nuclei4 26.4996140 0.8310685
Cell.size^8 25.6821554 0.8054318
Bare.nuclei3 24.5833551 0.7709717
Bare.nuclei7 21.8755488 0.6860507
Cell.size.C 20.0900403 0.6300545
Bare.nuclei5 19.8094615 0.6212551
Bl.cromatin7 18.8161804 0.5901043
Cl.thickness.Q 17.8206566 0.5588831
Epith.c.size.L 17.2140778 0.5398599
Cell.shape.C 15.8214699 0.4961855
Bare.nuclei8 15.7832696 0.4949875
Cell.shape^8 15.5873838 0.4888443
Normal.nucleoli6 15.5750643 0.4884579
Epith.c.size^8 15.1072819 0.4737875
Cell.size^5 14.7341024 0.4620840
Mitoses10 14.6725910 0.4601549
Cell.size.Q 13.5285932 -0.4242774
Cl.thickness^5 12.5086592 0.3922907
Normal.nucleoli7 11.2372547 -0.3524175
Bl.cromatin4 11.0371135 0.3461408
Epith.c.size^5 10.1426736 -0.3180898
Bl.cromatin8 8.5926819 0.2694797
Epith.c.size^9 8.1088768 0.2543068
Normal.nucleoli3 6.1834552 0.1939226
Cell.size^6 6.0301535 -0.1891149
Marg.adhesion.C 5.6067864 0.1758374
Marg.adhesion^8 5.0073879 -0.1570394
Epith.c.size^7 4.8165935 -0.1510558
Bl.cromatin10 3.7579941 0.1178565
Marg.adhesion^4 0.3997860 -0.0125379
Cell.shape^5 0.2113266 -0.0066275
Cl.thickness.C 0.1668182 0.0052317
(Intercept) NA 0.5035466