Automatic constraints generation for semisupervised clustering: experiences with documents classification

In the last times, semi-supervised clustering has been an area that has received a lot of attention. It is distinguished from more traditional unsupervised approaches on the use of a small amount of supervision to steer clustering. Unfortunately in the real world, the supervision is not always available: data to process are often too large and so the cost (in terms of time and human resources) for user-provided information is not conceivable. To address this issue, this work presents an automatic generation of the supervision, by the analysis of the data structure itself. This analysis is performed using a partitional clustering algorithm that discovers relationships between pairs of instances that may be used as a semi-supervision in the clustering process. The methodology has been studied in the document clustering domain, an area where novel approaches for accurate documents classifications are strongly required. Experimental result shows the validity of this approach.