Reference based objects • compboost

In order to provide high flexibility, we decided to write the base implementation of compboost in C++ to make use of object oriented programming. These C++ classes can be used in R after exposing them as S4 classes. To take away this layer of abstraction we decided to break these S4 classes to one R6 wrapper class.

All in all, the class system of compboost is a mixture of raw exposed S4 classes and the convenience class written in R6. As usual for object oriented programming, the classes works on a reference base. This introduction aims to show how these references work and can be accessed.

The main classes that is most affected by references is the Response class.

Response Class

The target variable is represented as an object that inherits from the Response class. Depending on the target type we like to have different transformations of the internally predicted scores. For instance, having a binary classification task the score $\hat{f}(x) \in \mathbb{R}$ is transformed to a $[0,1]$ scale by using the logistic function:

\[ \hat{\pi}(x) = \frac{1}{1 + \exp(-\hat{f}(x))} \]

To show the how references work here, we first define a ResponseBinaryClassif object. Therefore, we use the mtcars dataset and create a new binary target variable for fast $\text{qsec} < 17$ or slow $\text{qsec} \geq 17$ cars and create the response object:

df = mtcars[, c("mpg", "disp", "hp", "drat", "wt")]
df$qsec_cat = ifelse(mtcars$qsec < 17, "fast", "slow")

obj_response = ResponseBinaryClassif$new("qsec_cat", "fast", df$qsec_cat)
obj_response
#> 
#> Binary classification response of target "qsec_cat" and threshold 0.5
#> ResponseBinaryClassifPrinter

To access the underlying representation of the response class (here a binary variable) one can use $getResponse(). In the initialization of a new response object, the prediction $\hat{f} \in \mathbb{R}$ is initialized with zeros. We can also use the response object to calculate the transformed predictions $\hat{\pi} \in [0,1]$:

knitr::kable(head(data.frame(
  target = df$qsec_cat,
  target_representation = obj_response$getResponse(),
  prediction_initialization = obj_response$getPrediction(),
  prediction_transformed = obj_response$getPredictionTransform()
)))

target	target_representation	prediction_transformed
fast	1	0.5
slow	-1	0.5
slow	-1	0.5
slow	-1	0.5
slow	-1	0.5
slow	-1	0.5

In the case of binary classification, we can use the response object to calculate the predictions on a label basis by using a specified threshold $a$: \[ \hat{y} = 1 \ \ \text{if} \ \ \hat{\pi}(x) \geq a \]

The default threshold here is 0.5:

obj_response$getThreshold()
#> [1] 0.5
head(obj_response$getPredictionResponse())
#>      [,1]
#> [1,]    1
#> [2,]    1
#> [3,]    1
#> [4,]    1
#> [5,]    1
#> [6,]    1

By setting the threshold to 0.6 we observe now that each class is predicted as negative:

obj_response$setThreshold(0.6)
head(obj_response$getPredictionResponse())
#>      [,1]
#> [1,]   -1
#> [2,]   -1
#> [3,]   -1
#> [4,]   -1
#> [5,]   -1
#> [6,]   -1

This behavior has nothing to do with references at the moment. During the fitting of a component-wise boosting model, these predictions are adjusted over and over again by the Compboost object. This is where the reference comes in:

cboost = boostSplines(data = df, target = obj_response,
  iterations = 2000L, trace = 500L)
#>    1/2000   risk = 0.59  time = 0   
#>  500/2000   risk = 0.22  time = 16384   
#> 1000/2000   risk = 0.15  time = 44252   
#> 1500/2000   risk = 0.12  time = 82352   
#> 2000/2000   risk = 0.1  time = 129766   
#> 
#> 
#> Train 2000 iterations in 0 Seconds.
#> Final risk based on the train set: 0.1

Having again a look at the predictions shows the difference to the values before training. During the fitting process, the predictions of the response object are updated by the model:

knitr::kable(head(data.frame(
  target = df$qsec_cat,
  prediction = obj_response$getPrediction(),
  prediction_transformed = obj_response$getPredictionTransform(),
  prediction_response = obj_response$getPredictionResponse()
)))

target	prediction	prediction_transformed	prediction_response
fast	-0.0082833	0.4979292	-1
slow	-1.0690678	0.2555804	-1
slow	-3.1015214	0.0430445	-1
slow	-5.9979559	0.0024777	-1
slow	-2.2784370	0.0929246	-1
slow	-2.7290270	0.0612821	-1