On Reproducible Implementations in Unsupervised Concept Drift Detection Algorithms Research

Abstract

In order to create reproducible experimentation and algorithms in machine learning and data mining research, reproducible descriptions of the algorithms are needed. These can be in the form of source code, pseudo code and prose. Efforts in academia commonly focus on accessibility of source code. Based on an internal study reproducing unsupervised concept drift detectors, this work argues that a publication’s content is equally important and highlights common issues affecting attempts at implementing unsupervised concept drift detectors. These include major issues prohibiting implementation entirely, as well as minor issues, which demand increased effort from the developer. The paper proposes the use of a checklist as a consistent tool to ensure better quality and reproducible publications of algorithms. The issues highlighted in this work could mark a starting point, although future work is required to ensure representation of more diverse areas of research in artificial intelligence.