Oosterhuis, J. (2020). Weight learning in LP MLN for collective classification [Diploma Thesis, Technische Universität Wien]. reposiTUm. https://doi.org/10.34726/hss.2020.66402
In this thesis we investigate in detail the theory and practice of weight learning in LP MLN for collective classification problems. We test different weight learning methods on a practical collective classification problem, namely object labeling in images using relational context constraints, based on an exposition of the theory of weight learning in LP MLN. In our experiments we consider the applicability of different systems for weight learning in LP MLN and evaluate the performance of different weight learning methods in different scenarios. As we show, the best learning method depends very much on the input program and the specific dataset, and no single method noticeably outperforms other methods in our tests, with very simple learning methods performing very well. The performance of different methods can be very hard to predict, as some of our results are very unexpected. Nonetheless, from our experiments we conclude that weight learning noticeably improves the performance of an LP MLN-program in a collective classification setup and that effective weight learning methods and systems exist for LP MLN .
