pusion.core.cosine_similarity_combiner

class pusion.core.cosine_similarity_combiner.CosineSimilarityCombiner

Bases: pusion.core.combiner.UtilityBasedCombiner

The CosineSimilarityCombiner considers the classification assignments to \(\ell\) classes as vectors from an \(\ell\)-dimensional vector space. The normalized cosine-similarity measure between two vectors \(x\) and \(y\) is calculated as

\[cos(x,y) = \dfrac{x\cdot y}{|x||y|}\ .\]

The cosine-similarity is calculated pairwise and accumulated for each classifier for one specific sample. The fusion is represented by a classifier which shows the most similar classification output to the output of all competing classifiers.

combine(decision_tensor)

Combine decision outputs with as an output that accommodates the highest cosine-similarity to the output of all competing classifiers. In other words, the best representative classification output among the others is selected according to the highest cumulative cosine-similarity. This method supports both, continuous and crisp classification outputs.

Parameters

decision_tensor – numpy.array of shape (n_classifiers, n_samples, n_classes). Tensor of either crisp or continuous decision outputs by different classifiers per sample.

Returns

A matrix (numpy.array) of either crisp or continuous class assignments which represents fused decisions obtained by the highest cumulative cosine-similarity. Axis 0 represents samples and axis 1 the class labels which are aligned with axis 2 in decision_tensor input tensor.

class pusion.core.cosine_similarity_combiner.CRCosineSimilarity

Bases: pusion.core.cosine_similarity_combiner.CosineSimilarityCombiner

The CRCosineSimilarity is a modification of CosineSimilarityCombiner that also supports complementary-redundant decision outputs. Therefore the input is transformed, such that all missing classification assignments are considered as 0, respectively. To call combine() a coverage needs to be set first by the inherited set_coverage() method.

combine(decision_outputs)

Combine complementary-redundant decision outputs with as an output that accommodates the highest cosine-similarity to the output of all competing classifiers. In other words, the best representative classification output among the others is selected according to the highest cumulative cosine-similarity. This method supports both, continuous and crisp classification outputs.

Parameters

decision_outputs – list of numpy.array matrices, each of shape (n_samples, n_classes’), where n_classes’ is classifier-specific and described by the coverage. Each matrix corresponds to one of n_classifiers classifiers and contains crisp or continuous decision outputs per sample.

Returns

A matrix (numpy.array) of crisp or continuous class assignments which represents fused decisions. Axis 0 represents samples and axis 1 the class labels which are aligned with axis 2 in decision_tensor input tensor.