The Data Problem in Data Mining

Abstract

Computer science is essentially an applied or engineering science, creating tools. In Data Mining, those tools are supposed to help humans understand large amounts of data. In this position paper, I argue that for all the progress that has been made in Data Mining, in particular Pattern Mining, we are lacking insight into three key aspects: 1) How pattern mining algorithms perform quantitatively, 2) How to choose parameter settings, and 3) How to relate found patterns to the processes that generated the data. I illustrate the issue by surveying existing work in light of these concerns and pointing to the (relatively few) papers that have attempted to fill in the gaps. I argue further that progress regarding those questions is held back by a lack of data with varying, controlled properties, and that this lack is unlikely to be remedied by the ever increasing collection of real-life data. Instead, I am convinced that we will need to make a science of digital data generation, and use it to develop guidance to data practitioners.