Authors: Greg Ceccarelli, Michael Marks, Rock Baek
License: MIT
This vignette is designed to introduce you to the autoCaret R package. This package is built on top of both the caret and caretEnsemble R packages for machine learning and can take as input an R dataframe suitable for binary classification. Currently, binary classification, is the primary purpose of autoCaret.
The main function, automodel, will do the following:
- Validate the input data frame - binarizing the target variable if possible.
- Split the dataframe into both a training and test set
- Preprocess the input data - center, scale and remove variables with zero variance
- Check for class imbalance and attempt to downsample, as default, to help combat poor predictive accuracy
- Build a suite of models determined by the
methodlistparameter. If NULL, defaults to "glm", "rpart", rf" &"xgbLinear"" - Blend the models into an ensemble model using the
caretEnsemblepackage - Use the ensemble to make predictions on the held out test set
- Return, to the end user, an
autoCaretmodel object with a variety of useful pieces of metadata about the modeling process
The autoCaret model object is fully accessible and can be summarized using the the summary() generic function.
Additionally, predictions can be made using the predict() generic function passing in as parameters an autoCaret object and new data.
At the time of this writing, this package is not hosted on CRAN, but can be obtained from GitHub. To do so, first make sure you have devtools installed.
install.packages("devtools")Now we can install from GitHub using the following line:
devtools::install_github("gregce/autoCaret")Once the autoCaret package is installed, you may access its functionality as you would any other package by calling:
library("autoCaret")If all went well, check out the vignette("autoCaret") which will pull up this vignette!
We begin by loading the mlbench pakacage and some example data, Sonar, which is commonly used to demostrate machine learning functionality. In this example, we will attempt to distinguish Mines (M) from Rocks (R) using binary classification with an initial dataset of where N=208 and P=60.
As a general rule, when using autoCaret::autoModel defaults, datasets less than 100mb should yield optimal performance. and in order to avoid extremely long run times and/or high memory requirements.
library(mlbench)
library(autoCaret)
# Load the data into Memory from the mlbench package
data("Sonar")
# Take a brief peak at the Sonar dataframe
dplyr::glimpse(Sonar)## Observations: 208
## Variables: 61
## $ V1 <dbl> 0.0200, 0.0453, 0.0262, 0.0100, 0.0762, 0.0286, 0.0317, ...
## $ V2 <dbl> 0.0371, 0.0523, 0.0582, 0.0171, 0.0666, 0.0453, 0.0956, ...
## $ V3 <dbl> 0.0428, 0.0843, 0.1099, 0.0623, 0.0481, 0.0277, 0.1321, ...
## $ V4 <dbl> 0.0207, 0.0689, 0.1083, 0.0205, 0.0394, 0.0174, 0.1408, ...
## $ V5 <dbl> 0.0954, 0.1183, 0.0974, 0.0205, 0.0590, 0.0384, 0.1674, ...
## $ V6 <dbl> 0.0986, 0.2583, 0.2280, 0.0368, 0.0649, 0.0990, 0.1710, ...
## $ V7 <dbl> 0.1539, 0.2156, 0.2431, 0.1098, 0.1209, 0.1201, 0.0731, ...
## $ V8 <dbl> 0.1601, 0.3481, 0.3771, 0.1276, 0.2467, 0.1833, 0.1401, ...
## $ V9 <dbl> 0.3109, 0.3337, 0.5598, 0.0598, 0.3564, 0.2105, 0.2083, ...
## $ V10 <dbl> 0.2111, 0.2872, 0.6194, 0.1264, 0.4459, 0.3039, 0.3513, ...
## $ V11 <dbl> 0.1609, 0.4918, 0.6333, 0.0881, 0.4152, 0.2988, 0.1786, ...
## $ V12 <dbl> 0.1582, 0.6552, 0.7060, 0.1992, 0.3952, 0.4250, 0.0658, ...
## $ V13 <dbl> 0.2238, 0.6919, 0.5544, 0.0184, 0.4256, 0.6343, 0.0513, ...
## $ V14 <dbl> 0.0645, 0.7797, 0.5320, 0.2261, 0.4135, 0.8198, 0.3752, ...
## $ V15 <dbl> 0.0660, 0.7464, 0.6479, 0.1729, 0.4528, 1.0000, 0.5419, ...
## $ V16 <dbl> 0.2273, 0.9444, 0.6931, 0.2131, 0.5326, 0.9988, 0.5440, ...
## $ V17 <dbl> 0.3100, 1.0000, 0.6759, 0.0693, 0.7306, 0.9508, 0.5150, ...
## $ V18 <dbl> 0.2999, 0.8874, 0.7551, 0.2281, 0.6193, 0.9025, 0.4262, ...
## $ V19 <dbl> 0.5078, 0.8024, 0.8929, 0.4060, 0.2032, 0.7234, 0.2024, ...
## $ V20 <dbl> 0.4797, 0.7818, 0.8619, 0.3973, 0.4636, 0.5122, 0.4233, ...
## $ V21 <dbl> 0.5783, 0.5212, 0.7974, 0.2741, 0.4148, 0.2074, 0.7723, ...
## $ V22 <dbl> 0.5071, 0.4052, 0.6737, 0.3690, 0.4292, 0.3985, 0.9735, ...
## $ V23 <dbl> 0.4328, 0.3957, 0.4293, 0.5556, 0.5730, 0.5890, 0.9390, ...
## $ V24 <dbl> 0.5550, 0.3914, 0.3648, 0.4846, 0.5399, 0.2872, 0.5559, ...
## $ V25 <dbl> 0.6711, 0.3250, 0.5331, 0.3140, 0.3161, 0.2043, 0.5268, ...
## $ V26 <dbl> 0.6415, 0.3200, 0.2413, 0.5334, 0.2285, 0.5782, 0.6826, ...
## $ V27 <dbl> 0.7104, 0.3271, 0.5070, 0.5256, 0.6995, 0.5389, 0.5713, ...
## $ V28 <dbl> 0.8080, 0.2767, 0.8533, 0.2520, 1.0000, 0.3750, 0.5429, ...
## $ V29 <dbl> 0.6791, 0.4423, 0.6036, 0.2090, 0.7262, 0.3411, 0.2177, ...
## $ V30 <dbl> 0.3857, 0.2028, 0.8514, 0.3559, 0.4724, 0.5067, 0.2149, ...
## $ V31 <dbl> 0.1307, 0.3788, 0.8512, 0.6260, 0.5103, 0.5580, 0.5811, ...
## $ V32 <dbl> 0.2604, 0.2947, 0.5045, 0.7340, 0.5459, 0.4778, 0.6323, ...
## $ V33 <dbl> 0.5121, 0.1984, 0.1862, 0.6120, 0.2881, 0.3299, 0.2965, ...
## $ V34 <dbl> 0.7547, 0.2341, 0.2709, 0.3497, 0.0981, 0.2198, 0.1873, ...
## $ V35 <dbl> 0.8537, 0.1306, 0.4232, 0.3953, 0.1951, 0.1407, 0.2969, ...
## $ V36 <dbl> 0.8507, 0.4182, 0.3043, 0.3012, 0.4181, 0.2856, 0.5163, ...
## $ V37 <dbl> 0.6692, 0.3835, 0.6116, 0.5408, 0.4604, 0.3807, 0.6153, ...
## $ V38 <dbl> 0.6097, 0.1057, 0.6756, 0.8814, 0.3217, 0.4158, 0.4283, ...
## $ V39 <dbl> 0.4943, 0.1840, 0.5375, 0.9857, 0.2828, 0.4054, 0.5479, ...
## $ V40 <dbl> 0.2744, 0.1970, 0.4719, 0.9167, 0.2430, 0.3296, 0.6133, ...
## $ V41 <dbl> 0.0510, 0.1674, 0.4647, 0.6121, 0.1979, 0.2707, 0.5017, ...
## $ V42 <dbl> 0.2834, 0.0583, 0.2587, 0.5006, 0.2444, 0.2650, 0.2377, ...
## $ V43 <dbl> 0.2825, 0.1401, 0.2129, 0.3210, 0.1847, 0.0723, 0.1957, ...
## $ V44 <dbl> 0.4256, 0.1628, 0.2222, 0.3202, 0.0841, 0.1238, 0.1749, ...
## $ V45 <dbl> 0.2641, 0.0621, 0.2111, 0.4295, 0.0692, 0.1192, 0.1304, ...
## $ V46 <dbl> 0.1386, 0.0203, 0.0176, 0.3654, 0.0528, 0.1089, 0.0597, ...
## $ V47 <dbl> 0.1051, 0.0530, 0.1348, 0.2655, 0.0357, 0.0623, 0.1124, ...
## $ V48 <dbl> 0.1343, 0.0742, 0.0744, 0.1576, 0.0085, 0.0494, 0.1047, ...
## $ V49 <dbl> 0.0383, 0.0409, 0.0130, 0.0681, 0.0230, 0.0264, 0.0507, ...
## $ V50 <dbl> 0.0324, 0.0061, 0.0106, 0.0294, 0.0046, 0.0081, 0.0159, ...
## $ V51 <dbl> 0.0232, 0.0125, 0.0033, 0.0241, 0.0156, 0.0104, 0.0195, ...
## $ V52 <dbl> 0.0027, 0.0084, 0.0232, 0.0121, 0.0031, 0.0045, 0.0201, ...
## $ V53 <dbl> 0.0065, 0.0089, 0.0166, 0.0036, 0.0054, 0.0014, 0.0248, ...
## $ V54 <dbl> 0.0159, 0.0048, 0.0095, 0.0150, 0.0105, 0.0038, 0.0131, ...
## $ V55 <dbl> 0.0072, 0.0094, 0.0180, 0.0085, 0.0110, 0.0013, 0.0070, ...
## $ V56 <dbl> 0.0167, 0.0191, 0.0244, 0.0073, 0.0015, 0.0089, 0.0138, ...
## $ V57 <dbl> 0.0180, 0.0140, 0.0316, 0.0050, 0.0072, 0.0057, 0.0092, ...
## $ V58 <dbl> 0.0084, 0.0049, 0.0164, 0.0044, 0.0048, 0.0027, 0.0143, ...
## $ V59 <dbl> 0.0090, 0.0052, 0.0095, 0.0040, 0.0107, 0.0051, 0.0036, ...
## $ V60 <dbl> 0.0032, 0.0044, 0.0078, 0.0117, 0.0094, 0.0062, 0.0103, ...
## $ Class <fctr> R, R, R, R, R, R, R, R, R, R, R, R, R, R, R, R, R, R, R...
Having both the data loaded and having inspected it, we can now make use of the autoCaret::autoModel() function As stated above, we intend to try and distinguish Rocks (R) from Mines (R), so we will attempt to predict the Class variable in the Sonar dataframe.
Using it's defaults, autoModel has 2 arguments we need to specify: df and y.
df is the Dataframe that we'd like to use build a binary classification model, while y is our classification target or response variable. We can use a non-exported package function, autoCaret:::checkBinaryTrait to determine if our y variable is indeed binary. The autoModel functionality will perform this for us as well.
# Manually check that our intended y paramter is indeed binary
autoCaret:::checkBinaryTrait(Sonar$Class)## [1] TRUE
# Generate an autoCaret object using the autoModel function
mod <- autoCaret::autoModel(df = Sonar, y = Class, progressBar = FALSE)In the example above, the returned object, mod, is an autoCaret object containing 16 objects. To confirm, we can run the below two commmands:
# Check class of autoCaret object
class(mod)## [1] "autoCaret"
# High level
nrow(summary.default(mod))## [1] 19
Running the summary function on our model output displays a wealth of information about the contents of the object as well as the procedural steps taken during modeling. In our example, we observe:
- that our initial dataset of 208 observation was split into a training and test set containing 167 and 41 observations
- Modeling took .64 minutes and entailed resampling our dataset 10 times
- We used the four default models to create an ensemble.
- Using the ensemble model that was generated to predict on the test set yield predictions with 92% accuracy.
# Use the summary generic to store a summary of autoCaret object
overview <- summary(mod)## The input dataset had: 208 observations and 60 predictors
## ---------------------
## Prior to model training, the input dataset was split into a training & test set
## The training set has: 167 observations
## The test set has: 41 observations
## ---------------------
## Overall modeling took: 0.64 minutes
## During that time the training data was boostrap resampled 10 times
##
## The following classification models were used to create an ensemble:
## - glm
## - rpart
## - rf
## - xgbLinear
## To learn more about ensemble learing in the context of machine learning, please visit: https://en.wikipedia.org/wiki/Ensemble_learning
##
## In the ensemble, the top 5 variables in order from highest to lowest level of relative importance, were:
## - V11
## - V12
## - V9
## - V10
## - V36
## ---------------------
## When the ensemble model was used to predict on the held out test set of 41 observations it performed as follows:
##
## Overall Accuracy: 92.68
##
## A confusion matrix demonstrating accuracy is as follows:
## Reference
## Prediction R M
## R 18 2
## M 1 20
##
## Precision: 90
## Recall: 94.74
## ---------------------
## To learn more about Precision & Recall in the context of information retriveal, please visit: https://en.wikipedia.org/wiki/Precision_and_recall
We can also access each of object variables included in the above displayed summary output via the object itself.
# Print the overview to the console
overview## $input_row_count
## [1] 208
##
## $input_col_count
## [1] 61
##
## $train_row_count
## [1] 167
##
## $test_row_count
## [1] 41
##
## $modeling_time
## [1] 0.64
##
## $number_resamples
## [1] 10
##
## $method_list
## [1] "glm" "rpart" "rf" "xgbLinear"
##
## $confusionMatrix
## Reference
## Prediction R M
## R 18 2
## M 1 20
##
## $accuracy
## [1] 92.68
##
## $sensitivity
## [1] 94.74
##
## $specificity
## [1] 90.91
##
## $precision
## [1] 90
##
## $recall
## [1] 94.74
##
## $best_model_results
## model_name ROC Sens Spec ROCSD SensSD
## 1 ensemble 0.9337900 0.8095031 0.8953181 0.01498452 0.01089121
## 2 rf 0.9052501 0.6901687 0.8829980 0.03551813 0.10603065
## 3 xgbLinear 0.8702916 0.7168320 0.8185311 0.04529174 0.10874655
## 4 rpart 0.7019527 0.6562942 0.7122567 0.08602055 0.15284623
## 5 glm 0.6451431 0.5910052 0.6543618 0.08914517 0.08671003
## SpecSD
## 1 0.04371527
## 2 0.06784535
## 3 0.08614794
## 4 0.07226794
## 5 0.10772096
##
## attr(,"class")
## [1] "summary.autoCaret"
So now that we have a sense of how successful our auto modeling approach was, we'd likely want to use the model, mod, we built previously to make predictions on new data we receive.
Because this is an illustrative example, we'lll take a shortcut by just resampling the same data that we used to train on. The main point here is that you can simply pass your autoCaret model object, mod, into the predict() function along with new observations to generate predictions.
new <- Sonar[sample(1:nrow(Sonar), 50, replace=TRUE),]
#Make predicitons
preds <- predict(mod, new)
#Print Predictions
preds## [1] M R M R R R R M R M R M M R M R R M M R M M M M R M M M R R M R R M M
## [36] R M M M R R M M M R M M M M R
## Levels: R M
How well did we do? Well a confusion matrix from the caret package can tell us!
- We only mispredicted one example, for overall accuracy of .98
- Note: we wouldn't expect this level of accuracy using real data given we resampled from our original training set.
## How well did we do?
caret::confusionMatrix(data = preds, reference = new$Class)## Confusion Matrix and Statistics
##
## Reference
## Prediction M R
## M 29 0
## R 1 20
##
## Accuracy : 0.98
## 95% CI : (0.8935, 0.9995)
## No Information Rate : 0.6
## P-Value [Acc > NIR] : 2.775e-10
##
## Kappa : 0.9587
## Mcnemar's Test P-Value : 1
##
## Sensitivity : 0.9667
## Specificity : 1.0000
## Pos Pred Value : 1.0000
## Neg Pred Value : 0.9524
## Prevalence : 0.6000
## Detection Rate : 0.5800
## Detection Prevalence : 0.5800
## Balanced Accuracy : 0.9833
##
## 'Positive' Class : M
##