Redes neurais em grafos para aprendizado positivo: uma abordagem utilizando reescrita

Abstract

Given the increasing volume of data generated by society, the task of classification has become even more complex, as some supervised classification algorithms require a substantial amount of labeled data to achieve good performance. On the other hand, even with the use of a few labeled examples from the positive class, relevant results can be obtained by semi-supervised learning algorithms based on Positive and Unlabeled Learning (PUL). These algorithms can be applied to a variety of data types, including relational data, which can be represented by graphs. In this context, machine learning algorithms based on neural networks show relevant results, particularly Graph Neural Networks (GNNs). Additionally, the recent research area of graph rewiring has shown a performance gain in GNNs. This work aims to explore the use of graph rewiring and graph neural networks to address PUL problems, under the hypothesis that these approaches are capable of learning a latent representation space biased by the positive data. The results of experiments conducted on a range of datasets show that the proposed approach is superior to classical and current algorithms in the literature, both in negative data extraction and in the PUL task.