Local Embedding Consensus Across Multiple Descriptor Spaces

In many applications, real world objects may be described from different perspectives by different sets of attributes. This leads to multiple descriptor spaces in which the objects may be clustered in each of these spaces separately – but the clusters from different spaces do not necessarily match. In this paper we propose an approach to assess whether the objects are embedded in the respective spaces in a similar way, that is, if there is consensus about the local data neighbourhoods, which may then hint at clusters across multiple spaces. The approach does not employ a particular clustering algorithm to keep the methodology free from biases. The evaluation on artificial and real data shows that the measured consensus corresponds well to the intuition and provides interpretable insights.