Aggregation

Prediction with an ensemble model requires some form of aggregation of the predictions of its base models. In this page we will discuss what is currently available in EnsembleSVM.

In particular, we currently offer the following predefined aggregation schemes:

The parameters for each scheme are highlighted in blue. Default settings exist for the majority voting and logistic regression schemes (namely alpha a vector of ones and b=0). For the SVM and linear SVM, our software can load a LIBSVM or LIBLINEAR model from a file.

Ensemble models can be reconfigured using the esvm-edit tool. This allows modification of the aggregation method or reparameterization, among other things.

If you are interested in implementing aggregation schemes, please have a look at the following wiki pages related to implementation:

1. BasicBlock: the implementation of an elementary function. Many are already provided, such as multiplication with a scalar, the logistic function and more.
2. MultistagePipe: the implementation of a sequence of elementary functions. All the aggregation methods listed above are MultistagePipe objects. The framework allows straightforward prototyping of novel MultistagePipe objects which can also be serialized and deserialized with minimal effort.