Rastro-DM: data mining with a trail - A methodology for documenting data mining projects and its application in the construction of a text classifier of documents associated with damages to the public treasury

Resumo

This paper proposes a methodology for documenting data mining (DM) projects, Rastro-DM (Trail Data Mining), with a focus not on the model that is generated, but on the processes behind its construction, in order to leave a trail (Rastro in Portuguese) of planned actions, training completed, results obtained, and lessons learned. The proposed practices are complementary to structuring methodologies of DM, such as CRISP-DM, which establish a methodological and paradigmatic framework for the DM process. The application of best practices and their benefits is illustrated in a project called “Cladop” that was created for the classification of PDF documents associated with the investigative process of damages to the Brazilian Federal Public Treasury. Building the Rastro-DM kit in the context of a project is a small step that can lead to an institutional leap to be achieved by sharing and using the trail across the enterprise.