| Abstract: |  This thesis aims at efficient construction of relational features in the context of inductive logic programming. In order to speed up construction of relational features, we devise and implement two algorithms for theta-subsumption and two algorithms for propositionalization for so called hierarchical features. The algorithms for theta-subsumption called ReSumEr and ReCovEr are tested on real-life data. They are shown to be faster than state-of-the-art theta-subsumption algorithm Django in many, but not all cases. The propositionalization algorithms called RelF and HiFi are based on a theoretical analysis, which shows that relevancy and irreducibility properties of hierarchical features are monotone in a certain sense, which enables the algorithms to considerably reduce numbers of features, which need to be constructed. On three real-life datasets, RelF and HiFi are shown to construct features of lengths achievable neither by state-of-the-art propositionalization system RSD, nor by a general ILP system Progol. Predictive accuracies obtained on these datasets are close to best results reported in literature. Finally, complexity of feature construction is analyzed.
|
|---|