Learning qualitative dependencies

Jure Žabkar (2010) Learning qualitative dependencies. PhD thesis.

PDF Download (5Mb)

Abstract

The thesis presents novel approaches to learning qualitative models from given data. Learning qualitative models lies in the intersection of qualitative reasoning and machine learning. Qualitative models are abstractions and simplifications of numerical models. Qualitative modelling thus tends to learn simple and comprehensible models. Qualitative models are useful because they support human way of reasoning. Automatic learning of such models from data follows the main goal of artificial intelligence which is to make machines reason like humans. Another useful aspect of automatic learning of qualitative models is to make software tools that assist humans in understanding and discovering new knowledge from data. The motivation behind the methods that we develop in this thesis comes from the fact that, when inspecting the relation between two quantities, people most often consider only two quantities at a time. Although they do not make it explicit, they assume other quantities in the context constant. In mathematics, this principle is known as partial derivative and has been, since its discovery by Newton and Leibniz at the end of 17th century, an indispensable tool in mathematics and physics. The core of this thesis deals with the algorithms for computation of partial derivatives from data and learning qualitative models from partial derivatives. Taking mathematical definition of partial derivative as a foundation, we have developed six methods (with a common name Padé) for computation of partial derivatives in regression domains and a method (Qube) for computation of probabilistic partial derivatives in classification. We proposed a novel two-phase method for learning qualitative models from precomputed partial derivatives. In the first phase, we compute qualitative partial derivatives for each learning example and in the second phase we use an appropriate machine learning algorithm for classification to induce a qualitative model. As a part of our research, we have also developed four other algorithms for learning qualitative models and Q2 learning. We shortly describe these algorithms at the end of the thesis. We have implemented the above mentioned methods in Orange, an open source machine learning framework. We tested and evaluated them in controlled environment, a set of artificial domains, where we studied their properties. Further, we demonstrated how qualitative partial derivatives can be used in knowledge discovery from data. We conclude the experimental section with four realistic domains - two smaller robotic domains and two case studies. We used Padé in realistically simulated domain billiards and Qube in a real medical domain from infectology. In both cases we asked domain experts for explaining the induced qualitative models. Both algorithms induced simple, accurate and comprehensible models and proved useful in the conceptualization of the domains. We conclude the thesis with a short discussion of the results and possible further work.

Actions (login required)