Argument-based machine learning with logistic regression

David Možina (2017) Argument-based machine learning with logistic regression. MSc thesis.

Preview

PDF Download (1323Kb)

Abstract

People nowadays tend to use simple tools and procedures to process and analyze new data. We are eager to find easy solutions to extract useful information from it and build predictive models. In this thesis we designed a tool, which can easily be used to cope with a new data and which also enables a possibility to articulate expert's domain knowledge in the form of new features, i.e. attributes. The tool is based on a paradigm of argument-based machine learning (ABML) and machine learning method, called logistic regression. We modified the logistic regression method by adding the possibility to articulate the expert's domain knowledge and developed a new method, that allows interaction between domain expert and logistic regression called argument-based machine learning with logistic regression. We created an application with a graphical user interface that uses newly created method and, by using interactive loop, captures domain expert's knowledge. The knowledge is passed into the predictive model in the form of new attributes. Method searches for problematic examples which are examples that are wrongly predicted by a logistic regression model. These examples are presented to the domain expert. Expert's task is now to explain critical examples by giving arguments and providing explanations for wrongly predicted example. According to the given expert's arguments, method finds relevant couterexamples which can highlight possible flaws and shortcomings in the expert's arguments. Counterexamples change regulary, based on the conditions mentioned by the expert. The interaction between the expert and argument-based machine learning method can lead to better and more accurate models that are consistent with expert's domain knowledge. The newly created application also enables creating new attributes, which can be made during the argumentation process. If the newly created attribute solves current critical example, it gets replaced by a new problematic example. This leads to a faster interaction between a domain expert and the machine learning algorithm.

Actions (login required)