Copy the page URI to the clipboard
Tiddi, Ilaria; d'Aquin, Mathieu and Motta, Enrico
(2014).
DOI: https://doi.org/10.1007/978-3-319-13704-9_41
Abstract
Two typical problems are encountered after obtaining a set of rules from a data mining process: (i) their number can be extremely large and (ii) not all of them are interesting to be considered. Both manual and automatic strategies trying to overcome those problems have to deal with technical issues such as time costs and computational complexity. This work is an attempt to address the quantity and quality issues through using a Neural Network model for predicting the quality of Linked Data rules. Our motivation comes from our previous work, in which we obtained large sets of atomic rules through an inductive logic inspired process traversing Linked Data. Assuming a limited amount of resources, and therefore the impossibility of trying every possible combination to obtain a better rule representing a subset of items, the major issue becomes detecting the combinations that will produce the best rule in the shortest time. Therefore, we propose to use a Neural Network to learn directly from the rules how to recognise a promising aggregation. Our experiments show that including a Neural Network-based prediction model in a rule aggregation process significantly reduces the amount of resources (time and space) required to produce high-quality rules.