From machine learning towards knowledge discovery in databases

Abstract

The concept of knowledge discovery in databases (KDD) is an important future challenge for the different communities of machine learning (ML), statistics and databases. Although the aim of this new research area seems to be identified (the extraction of implicit, previously unknown, and potentially useful information from data), many criticise the lack of precision of this definition and, in particular, its similarity with the aim generally proclaimed for ML. We first recall the main purposes of ML and the proposed aims of KDD. That underlines the main new problems that KDD proposes to tackle which ML has not yet solved. Secondly, we specify the differences between both domains by answering specific comments that G. Piatetsky-Shapiro (1992, 1993) has made concerning this comparison of ML and KDD. It is important to clearly distinguish KDD from ML in order on the one hand to focus future ML research on the extension of current ML techniques aiming at larger KDD problems, and on the other hand for KDD researchers to know which ML techniques are best adapted to particular KDD tasks. We must, in the future, identify tasks that have not yet been sufficiently explored and then look for techniques developed in different scientific fields (in ML as well as in other fields) to solve comparable tasks.