New methods for discovering local behaviour in mixed databases

Abstract

Clustering techniques are widely used. There are many applications where it is desired to find automatically groups or hidden information in the data set. Finding a model of the system based in the integration of several local models is placed among other applications. Local model could have many structures; however, a linear structure is the most common one, due to its simplicity. This work aims at finding improvements in several fields, but all them will be applied to this finding of a set of local models in a database. On the one hand, a way of codifying the categorical information into numerical values has been designed, in order to apply a numerical algorithm to the whole data set. On the other hand, a cost index has been developed, which will be optimized globally, to find the parameters of the local clusters that best define the output of the process. Each of the techniques has been applied to several experiments and results show the improvements over the actual techniques.