Multi Class Datasets

We present 20 new multi-labeled artificial datasets, which can also be used for evaluating ambiguity resolving classifiers. The ambiguous or multi-labeled points are defined by those lying in the overlapping regions of two or more classes. Among the 20 datasets, 10 are 2-dimensional, while the rests are either 5 or 10-dimensional extended versions of the 2-dimensonal ones. The extensions are done following one of the two techniques. In the first strategy, datasets ate designed by appending 3 new dimensions each sampled uniformly at random and scaled between a specified range. The new 5-dimensional dataset is rotated by a random rotation matrix. This is a general technique by which any dataset can be transformed to higher dimensional feature space while conserving the properties of the ambiguous points. The second method extends the datasets by sampling them from a 10-dimensional real-valued feature space using the analogs class distributions of the corresponding 2-dimensional dataset. Such a strategy can extend a dataset to arbitrarily higher dimension feature space. However, the datasets will become sparse with increasing dimensionality. To tackle this issue the number of data points is increased in this case