Overview¶
Introduction¶
In general, there exist two main approaches to combine multiple classifiers in order to obtain reasonable classification outputs for unseen samples: the classifier selection and the classifier fusion. The latter one is followed by the presented framework and illustrated in Fig. 1.
Fig. 1 Architectural embedding of the decision fusion framework¶
However, there are significant properties arising from such classifier ensembles as parameters which need to be considered in the whole decision fusion process. These parameters are forming a configuration which consists of:
Decision fusion method (combiner)
Classification problem
Class assignment type
Classification coverage
The combiner states an explicit method provided by the framework which should be applied on the input data set. Available core methods are listed in the following section.
The classification problem refers either to a multiclass or to a multilabel classification problem. In the multiclass case, a sample is always classified into one class, while in the multilabel case, more than one class may be assigned to a sample.
Pusion operates on classification data which is given by class assignments. The class assignment describes memberships to each individual class for a sample. A class assignment type is either crisp or continuous. Crisp assignments are equivalent to labels and continuous assignments represent probabilities for each class being true.
The classification coverage states for each input classifier, which classes it is able to decide. A classifier ensemble may yield a redundant, complementary or complementary-redundant coverage.
Core methods¶
The following core decision fusion methods are supported by pusion and classified according to the evidence resolution they accept. With lower evidence resolution, a weaker a-priori information about individual classifiers can be taken into account during the fusion process. Utility-based methods do not take any further information about classifiers into account. In cases where no evidence is available for a certain classification data set, a utility-based method is a reasonable choice. Evidence-based methods are recommended in cases where evidence (e.g. confusion matrices) but no training data is available for each classifier. Trainable methods provide the highest evidence resolution, since decision outputs are required from the ensemble for each sample during the training phase. Therefore, each trainable combiner is able to calculate any kind of evidence based on the training data and is even able to analyse the behaviour of each classification method from the ensemble.
Utility-based methods (low evidence resolution):
Borda Count (BC)
Cosine Similarity (COS)
Macro Majority Vote (MAMV)
Micro Majority Vote (MIMV)
Simple Average (AVG)
Evidence-based methods (medium evidence resolution):
Trainable methods (highest evidence resolution):
Data input and output¶
The input type used for classification data is generic and applies to all provided decision fusion methods. It is given by a 3D numpy.ndarray tensor, which is illustrated in Fig. 2.
Fig. 2 Illustration of the input tensor for a multilabel problem with crisp assignments (3 samples, 4 classes, 2 classifiers).¶
The same applies also to the pusion’s return, except that the output matrix is a 2D numpy.ndarray.
Note
In case of complementary-redundant decisions, the coverage needs to be specified besides ordinary python lists, which are used as an alternative to the numpy.ndarray.
AutoFusion¶
The framework provides an additional fusion method AutoCombiner which is able to the detect the configuration based on the input classification data and to automatically select the fusion method with the best classification performance for the given problem. The AutoCombiner bundles all methods provided by the framework and probes each of them for the application on the given classification data. The AutoCombiner is transparent to the user as each of the core fusion methods.
Generic fusion¶
In contrast to the AutoCombiner, the GenericCombiner retrieves fusion results obtained by all compatible core methods by means of a numpy.ndarray tensor. In this case, the evaluation as well as the method selection is handed over to the user.
Further functionalities¶
Classification data and coverage generation (see module generator)
Useful transformations for decision outputs, e.g. multilabel to multiclass conversion (see module transformer)
Evaluation methods for different classification and coverage types (see class Evaluation)