Unsupervised Horizontal Collaboration Based in SOM

In this paper we present a new approach of collaborative classification allowing protecting the confidentiality of the data by using the self organizing map of Kohonen.

Having a collection of databases distributed on several different sites, so, the problem consists in clustering each of these bases by considering the data and the classifications of the others base coworkers, without omitting however to respect the constraint of confidentiality which forbids the sharing of data between the various centers. To do it, our approach is subdivided into two phases: a local phase and a collaborative phase.

The local phase would mean applying the classic algorithm of Kohonen, locally and independently on each of the databases, what will end in the obtaining of a map (SELF-ORGANIZING MAP) for each of these bases. The phase of collaboration would mean making each of the databases collaborate with all the map SOM partners in the other bases obtained during the local phase. So, as result we obtain on each of the sites a map close to the SOM which we would have obtained if we had disregarded the constraint of confidentiality, namely make databases collaborate they same. In the stemming from both phases, all the maps will be enriched. The article presents the formalism of the approach as well as its validation.

The proposed approach was validated on several databases and the experimental results showed very promising performances.