Relational approaches for learning, transferring and mining

Machine learning aims to design algorithms whose performance on a task improves with experience, where experience is usually defined as the amount of available data. Unfortunately, most traditional machine-learning algorithms rely on assumptions that do not hold in important real-world applications. One assumption is that the data have a simple structure, while most real-world problems involve complex, relational data. Another assumption is that large quantities of data are available to learn a sufficiently accurate predictive model, while high-quality data are often scarce in real-world domains. This dissertation aims to overcome the limitations of traditional machine-learning algorithms by proposing approaches for learning predictive models in domains that are characterized by a complex, relational structure and a shortage of high-quality data to learn accurate models. Furthermore, this dissertation also applies relational-learning techniques to spatio-temporal sports data, which are illustrative for the challenges that many other real-world applications pose. Most of the work in this dissertation relates to the field of statistical relational learning, which is concerned with learning predictive models for domains exhibiting both uncertainty and a complex structure. To address the inherent challenges, statistical-relational-learning formalisms typically combine a probabilistic model with a relational representation. This dissertation leverages Markov logic networks, which combine Markov random fields with logic. This dissertation presents five main contributions. The first contribution is an algorithm for learning Markov random fields from binary data. The second contribution is an algorithm for learning Markov logic networks from relational data. The third contribution is an algorithm for transferring knowledge across relational domains, where the domains can be entirely different. The fourth contribution is an approach for discovering offensive strategies in spatio-temporal soccer match data. The fifth contribution is an approach for discovering offensive patterns in spatio-temporal volleyball match data.

Jaar van publicatie:2016

Toegankelijkheid:Open