Relational learning algorithms combine machine learning and knowledge representation to learn from structured data and relationships between entities. Two prominent approaches are Statistical Relational Learning (SRL), which uses logical representations, and Graph Neural Networks (GNNs), which learn from graph-structured data. Although both are promising for handling more complex data, they share a problem with traditional approaches: when there is a distribution shift between the training and test sets, a new model must be trained on new data. Transfer learning addresses this limitation by leveraging knowledge from pre-trained models, enabling faster, less data-intensive adaptation to new tasks. However, most research focuses on what knowledge to transfer and how to transfer it, while from where to transfer remains largely unexplored. As a result, source–target domain pairs are often chosen ad hoc. In this thesis, we propose SMART-TL, which indicates beforehand from where to transfer, given a set of pre-trained source domains to learn a new target domain. It relies on the underlying probabilities of both domains and computes their divergences. To do this, it relies on propositionalization, a method for representing relational domains in a high-dimensional feature space to more easily compare distributions. To evaluate our proposed method, we analyze the performance of three state-of-the-art transfer learning-based algorithms, two SRL-based and one GNN-based, and compare data distributions using Kullback-Leibler (KL) divergence. Our results show that algorithms perform best when transferring between the most similar pairs of source and target domains. However, our method can be more time-consuming in the GNN context than the brute-force approach.
.png)
.png)
