Extracted information quality, a comparative study in high and low dimensions

Uncovering interesting groups in either multidimensional or network spaces has become an essential mechanism for data exploration and understanding. Decision making requires relevant information as well as high-quality on the retrieved conclusions. We presented a comparative study of two compact representations drawn from the same set of data objects by clustering high-dimensional spaces and low-dimensional Barnes-Hut t-stochastic neighbour embeddings. There is no consensus on how the problem should be addressed and how these representations/models should be analysed because of their different notions. We introduced a measure to compare their results and capability to provide insights into the information retrieved. We considered low-dimensional embeddings as a potentially revealing strategy to uncover dynamics possibly not uncovered in big-data spaces. We demonstrated that a non-guided approach can be as revealing as a user-guided approach for data exploration and presented coherent results for good uncertainty modelling capability in terms of fuzziness and densities.