Machine learning dissected

Abstract

The research presented in this thesis addresses machine learning techniques and their application in the context of classification problems. Furthermore as this thesis is centered around a medical initiative (Behapp) the insights found were applied to the data produced by this initiative.

The direction of study on general machine learning techniques was chosen in order to model the knowledge on how to create optimized machine learning models. Furthermore, since it concerns the analysis of a medical data set the usage of transparent modeling techniques is prefered allowing us to relate the input (data) to the output (classification). This relates back to the goal of creating optimized models since transparent techniques are known to be outperformed by their non transparent counterparts.

Using the modeling approach by Weerd and Brinkkemper (2008) the machine learning techniques were modelled into a method in the form of a process-deliverable-diagram. The method was then applied to two datasets to evaluate the potential for improvements in performance.

We found that models generated using our method showed increased performance in terms of classification accuracy and overall reliability of the results. Next we applied transparent modeling techniques and the sociability scoring model (Eskes et al., 2016) to the data of the Behapp initiative. As expected, the in-depth look reveals various patterns where patients and controls are separated in the data.

In light of the results we feel that the method created enables further reasoning on the application of machine learning techniques in a single procedural data mining approach and may be extended to include procedures relevant to other domains. Last we find that the concept of an aggregated sociability score shows promise in expressive value having applied it to patient data for the first time.