Clustered multidimensional scaling for exploration in information retrieval
Eniko-Melinda Szekely, Éric Bruno, Stéphane Marchand‐Maillet · Archive ouverte UNIGE (University of Geneva) · 2007
The data that needs to be processed nowadays is frequently represented in high-dimensional spaces with the dimension given by the number of features selected.There is a gap between human perception of low-dimensional spaces and the behaviour of distances within high-dimensional spaces.In data analysis the phenomenon of "curse of dimensionality" has consequences on the (dis)similarity matrices because the points become equidistant.In such a situation, methods for dimensionality reduction fail to reveal in the low-dimensional projected space structures existing in the data.We therefore propose in this article a clustered multidimensional scaling method for the discovery and understanding of data structures in view of exploration.Firstly, the data is clustered in the original space based on the closest k neighbours of each point which results in a disconnected graph.Secondly, an MDS is performed on each of the graph components.And finally the clusters' representatives are projected in the reduced space by means of an MDS in order to preserve the distances between clusters from the original space.