Exploratory and diagnostic machine learning tools for R
The goal of this package is multifold:
- Speed up data preparation for feeding machine-learning models
- Identify structure and patterns in a dataset
- Evaluate the results of a machine-learning model
install.packages("mltools")
install.packages("devtools")
devtools::install_github("ben519/mltools")
Predict whether or not someone is an alien.
library(data.table)
library(mltools)
# Copy the toy datasets since they are locked from being modified
train <- copy(alientrain)
test <- copy(alientest)
train
SkinColor IQScore Cat1 Cat2 Cat3 IsAlien
1: green 300 type1 type1 type4 TRUE
2: white 95 type1 type2 type4 FALSE
3: brown 105 type2 type6 type11 FALSE
4: white 250 type4 type5 type2 TRUE
5: blue 115 type2 type7 type11 TRUE
6: white 85 type4 type5 type2 FALSE
7: green 130 type1 type2 type4 TRUE
8: white 115 type1 type1 type4 FALSE
test
SkinColor IQScore Cat1 Cat2 Cat3
1: white 79 type4 type5 type2
2: green 100 type4 type5 type2
3: brown 125 type3 type9 type7
4: white 90 type1 type8 type4
5: red 115 type1 type2 type4
- Are there any pairs of categorical fields which are highly/perfectly correlated?
- Are there any parent-child related categorical fields?
- How does the target variable change with IQScore?
- What's the cardinality and skewness of each feature?
# Combine train (excluding IsAlien) and test
alien.all <- rbind(train[, !"IsAlien", with=FALSE], test)
#--------------------------------------------------
## Check for correlated and hierarchical fields
gini_impurities(alien.all, wide=TRUE) # weighted conditional gini impurities
Var1 Cat1 Cat2 Cat3 SkinColor
1: Cat1 0.0000000 0.3589744 0.0000000 0.4743590
2: Cat2 0.0000000 0.0000000 0.0000000 0.3461538
3: Cat3 0.0000000 0.3589744 0.0000000 0.4743590
4: SkinColor 0.4102564 0.5384615 0.4102564 0.0000000
# (Cat1, Cat3) = (Cat3, Cat1) = 0 => Cat1 and Cat3 perfectly correspond to each other
# (Cat1, Cat2) > 0 and (Cat2, Cat1) = 0 => Cat1-Cat2 exhibit a parent-child relationship.
# You can guess Cat1 by knowing Cat2, but not vice-versa.
#--------------------------------------------------
## Check relationship between IQScore and IsAlien by binning IQScore into groups
train[, BinIQScore := bin_data(IQScore, bins=seq(0, 300, by=50))]
IQScore BinIQScore
1: 300 [250, 300]
2: 95 [50, 100)
3: 105 [100, 150)
4: 250 [250, 300]
5: 115 [100, 150)
6: 85 [50, 100)
7: 130 [100, 150)
8: 115 [100, 150)
train[, list(Samples=.N, IQScore=mean(IQScore)), keyby=BinIQScore]
BinIQScore Samples IQScore
1: [50, 100) 2 90.00
2: [100, 150) 4 116.25
3: [250, 300] 2 275.00
# Remove column BinIQScore
train[, BinIQScore := NULL]
#--------------------------------------------------
## Check skewness of fields
skewness(alien.all)
$SkinColor
SkinColor Count Pcnt
1: white 6 0.46153846
2: green 3 0.23076923
3: brown 2 0.15384615
4: blue 1 0.07692308
5: red 1 0.07692308
$Cat1
Cat1 Count Pcnt
1: type1 6 0.46153846
2: type4 4 0.30769231
3: type2 2 0.15384615
4: type3 1 0.07692308
...
- Cateogrical fields in train and test should be factors with the same levels
- Split the training dataset to do cross validation
- Convert datasets to sparses matrices
set.seed(711)
#--------------------------------------------------
## Set SkinColor as a factor, such that it has the same levels in train and test
## Set low frequency skin colors (1 or fewer occurences) as "_other_"
skincolors <- list(train$SkinColor, test$SkinColor)
skincolors <- set_factor(skincolors, aggregationThreshold=1)
train[, SkinColor := skincolors[[1]] ] # update train with the new values
test[, SkinColor := skincolors[[2]] ] # update test with the new values
# Repeat the process above for other categorical fields (without setting low freq. values as "_other_")
for(col in c("Cat1", "Cat2", "Cat3")){
vals <- list(train[[col]], test[[col]])
vals <- set_factor(vals)
set(train, j=col, value=vals[[1]])
set(test, j=col, value=vals[[2]])
}
#--------------------------------------------------
## Randomly split the training data into 2 equally sized datasets
# Partition train into two folds, stratified by IsAlien
train[, FoldID := folds(IsAlien, nfolds=2, stratified=TRUE, seed=2016)]
cvtrain <- train[FoldID==1, !"FoldID"]
SkinColor IQScore Cat1 Cat2 Cat3 IsAlien
1: green 300 type1 type1 type4 TRUE
2: brown 105 type2 type6 type11 FALSE
3: green 130 type1 type2 type4 TRUE
4: white 115 type1 type1 type4 FALSE
cvtest <- train[FoldID==2, !"FoldID"]
SkinColor IQScore Cat1 Cat2 Cat3 IsAlien
1: white 95 type1 type2 type4 FALSE
2: white 250 type4 type5 type2 TRUE
3: _other_ 115 type2 type7 type11 TRUE
4: white 85 type4 type5 type2 FALSE
#--------------------------------------------------
## Convert cvtrain and cvtest to sparse matrices
## Note that unordered factors are one-hot-encoded
library(Matrix)
cvtrain.sparse <- sparsify(cvtrain)
4 x 21 sparse Matrix of class "dgCMatrix"
SkinColor__other_ SkinColor_brown SkinColor_green SkinColor_white IQScore Cat1_type1 ...
[1,] . . 1 . 300 1
[2,] . 1 . . 105 .
[3,] . . 1 . 130 1
[4,] . . . 1 115 1
cvtest.sparse <- sparsify(cvtest)
4 x 21 sparse Matrix of class "dgCMatrix"
SkinColor__other_ SkinColor_brown SkinColor_green SkinColor_white IQScore Cat1_type1 ...
[1,] . . . 1 95 1
[2,] . . . 1 250 .
[3,] 1 . . . 115 .
[4,] . . . 1 85 .
- What was the model's AUC ROC score?
- How good was the model's predictions for each sample?
#--------------------------------------------------
## Naive model that guesses someone is an alien if their IQScore is > 130
cvtest[, Prediction := ifelse(IQScore > 130, TRUE, FALSE)]
#--------------------------------------------------
## Evaluate predictions
# Area Under the ROC Curve (AUC ROC)
auc_roc(preds=cvtest$Prediction, actuals=cvtest$IsAlien)
0.75
# Individual scores to determine which predictions were good/bad (see help(roc_scores) for details)
cvtest[, ROCScore := roc_scores(preds=Prediction, actuals=IsAlien)]
cvtest[order(ROCScore)]
SkinColor IQScore Cat1 Cat2 Cat3 IsAlien Prediction ROCScore
1: white 95 type1 type2 type4 FALSE FALSE 0.0000000
2: white 250 type4 type5 type2 TRUE TRUE 0.0000000
3: white 85 type4 type5 type2 FALSE FALSE 0.0000000
4: _other_ 115 type2 type7 type11 TRUE FALSE 0.1666667
If you'd like to contact me regarding bugs, questions, or general consulting, feel free to drop me a line - bgorman519@gmail.com
Found this package helpful? Show your support and buy some merch!