Transfer learning by mapping and revising boosted relational dependency networks (Q2203326)

Transfer learning methods have been employed increasingly often, but most of them do not take into account relationships between entities. The paper under review addresses transfer learning in the context of the area known as statistical relational learning (SRL). It presents an algorithm called TreeBoostler, which transfers the SRL state-of-the-art Boosted Relational Dependency Networks learned in the source domain to the target domain. It is based on the predicate mapping algorithm proposed by \textit{L. Mihalkova} et al. to transfer clauses in Markov logic networks [``Mapping and revising Markov logic networks for transfer learning'', in: Proceedings of the 22nd national conference on artificial intelligence Vol. 1, AAAI'07. Palo Alto: AAAI Press. 608--614 (2007)], and consists of two phases: first, it transfers the source boosted trees structure to the target domain by finding an adequate predicate mapping, then it revises those trees by pruning and expanding nodes. The authors have experimentally validated that the revision phase improves the performance of the transfer process, though, at the same time, it increases its computational costs because it introduces another search. They have also compared the TreeBoostler on seven publicly available datasets with three baseline methods that learn from scratch. The results of those comparisons show that the TreeBoostler learns more accurate models than the baselines, no matter whether with minimal target data or with an increasing amount of them.