Transfer learning for boosted relational dependency networks through genetic algorithms

Abstract

Machine learning improves models with a set of observations, aiming to find regularities. However, traditional machine learning methods fail at finding patterns from several objects and their relationships. Statistical relational learning goes a step further to discover patterns from relational domains and deal with data under uncertainty. Most machine learning methods assume the training and test data come from the same distribution and feature space. Nonetheless, in several scenarios, this assumption does not hold. Transfer learning is a technique that leverages learned knowledge from a source task to improve the performance in a target task when data is scarce. A costly challenge associated with transfer learning in relational domains is mapping from the source and target vocabularies. This dissertation proposes GROOT, a framework that applies genetic algorithm and their variations to discover the best mapping between the source and target tasks and adapt the transferred model. GROOT relies on a set of relational regression trees built from the source data as a starting point to build the models for the target task. Over generations, individuals carry a possible mapping. They are submitted to genetic operators that recombine subtrees and revise the initial structure tree, enabling a prune or expansion of the branches. We also propose a new algorithm called modified Biased Random-Key genetic algorithm (mBRKGA), a BRKGA-based method and show the space complexity calculation of the proposed mapping in the framework. Experimental results conducted in real-world datasets show that GROOT reaches results, as AUC ROC, better than the baselines in most cases.