Fast and Flexible Component-Wise Boosting

Daniel Schalk

November 29, 2018


  • We own a small booth at the Christmas market that sells mulled wine.

  • As we are very interested in our customers’ health, we only sell to customers who we expect to drink less than 15 liters per season.

  • To estimate how much a customer drinks, we have collected data from 200 customers in recent years.

  • These data include mulled wine consumption (in liter and cups), age, sex, country of origin, weight, body size, and 200 characteristics gained from app usage (that have absolutely no influence).

mw_consumption mw_consumption_cups gender country age weight height app_usage1
12.6 42 f Seychelles 21 119.25 157.9 0.1680
2.1 7 f Poland 57 67.72 157.5 0.8075
12.6 42 m Seychelles 25 65.54 181.6 0.3849
9.3 31 f Germany 68 84.81 195.9 0.3277
4.8 16 m Ireland 39 85.30 163.9 0.6021
9.3 31 f Germany 19 102.75 173.8 0.6044
5.1 17 f Poland 66 98.54 199.7 0.1246
7.5 25 f Seychelles 62 108.21 198.1 0.2946
15.0 50 m Seychelles 22 62.08 180.1 0.5776
8.4 28 m Germany 52 75.42 170.0 0.6310

With these data we want to answer the following questions:

  • Which of the customers’ characteristics are important to be able to determine the consumption?
  • How does the effect of important features look like?
  • How does the model behave on unseen data?

What can we do to answer all the questions?

What can we do to answer all the questions?
  • Fit a linear model?
  • Fit a linear model on each feature and build the ensemble?
  • Fit a regularized linear model?
  • Train a random forest?

What can we do to answer all the questions?
  • Fit a linear model?
    \(\rightarrow\) Not possible since \(p > n\).

  • Fit a linear model on each feature and build the ensemble?
    \(\rightarrow\) Possible, but how should we determine important effects?

  • Fit a regularized linear model?
    \(\rightarrow\) Possible, but with the linear model we get just linear effects.

  • Train a random forest?
    \(\rightarrow\) Possible, but we want to interpret the effects.

Component-Wise Boosting

The Idea of Component-Wise Boosting

Why Component-Wise Boosting?

  • Inherent (unbiased) feature selection.

  • Resulting model is sparse since important effects are selected first and therefore it is able to learn in high-dimensional feature spaces (\(p \gg n\)).

  • Parameters are updated iteratively. Therefore, the whole trace of how the model evolves is available.

About compboost

The compboost package is a fast and flexible framework for model-based boosting completely written in C++:

  • With mboost as standard, we want to keep the modular principle of defining custom base-learner and losses.

  • Completely written in C++ and exposed by Rcpp to obtain high performance and full memory control.

  • R API is written in R6 to provide convenient wrapper.

  • Major parts of the compboost functionality are unit tested against mboost to ensure correctness.

Applying compboost to the Use-Case

Quick Start With Wrapper Functions

Effect Visualization

Inbag and OOB Behavior

To get an idea, how the model behaves on unseen data we use 75 % as training data and the other 25 % of the data to calculate the out of bag (OOB) risk:

Define Model and Base-Learner the “Object-Oriented Style”

OOB Data

To track the OOB risk we have to prepare the new data so that compboost knows the new data sources:

Define Logger

Extract Inbag and OOB Data

Visualization of Inbag and OOB Risk

Set Model to a Specific Iteration

Feature Importance

Using Custom Losses - Defining a Loss

As customers do not buy the wine in liters but per cup, it might be better to use a Poisson loss to take the counting data into account. For that reason we define a custom loss:

Using Custom Losses - Using the Loss

Further Functionalities

  • Each logger can also be used as stopper, therefore we can use them for early stopping

  • In combination with the custom loss, we can use the OOB logger to track performance measures like the AUC (in binary classification)

  • Losses and base-learner can also be directly extended using C++ (see getCustomCppExample())

From C++ to R


  • Automated conversion between R and C++ data structures, such as vectors, matrices, or even whole classes
  • Seamless integration of Armadillo for linear algebra
  • Complicated stuff like compilation, or again, the conversion between R and C++ are handled automatically

C++ to R Wrap Up of compboost

Challenges When Using C++ and Rcpp

  • Saving object is not possible at the moment
  • Memory managing is not easy \(\Rightarrow\) Segmentation folds or memory leaks may happen
  • Exported API of classes is not very informative
  • Debugging of C++ from R can be very annoying and time-consuming

What’s Next?

  • Better selection process of base-learner
  • Speed up the training by parallel computations
  • Greater functionality:
    • Functional data structures and loss functions
    • Unbiased feature selection
    • Effect decomposition into constant, linear, and non-linear

Thanks for your attention!

Build Status AppVeyor Build Status Coverage Status license CRAN_Status_Badge status

Actively developed on GitHub:

Project page:


Slides were created with: