Avaliação de métodos de construção de grafos para classificação no aprendizado semi-supervisionado

Abstract

Semi-supervised learning has gained relevance in classification tasks where only a fraction of the samples are labeled. In this context, graph-based methods have proven effective by leveraging the similarity structure among data points. This work investigates the impact of different graph construction strategies on the performance of label propagation, comparing four approaches: kNN (baseline), RGCLI, KAOG, and SNGC. The methods were evaluated on five datasets with varying proportions of labeled data and parameter settings. Results show that although the kNN method — which does not use supervision — achieved competitive performance in most scenarios, RGCLI stood out among the supervised methods, combining robustness with good generalization. It was also observed that the performance of the methods varies according to the complexity and structure of the datasets, with SNGC being more effective on high-dimensional data, and KAOG limited by its low connectivity. The statistical analysis confirmed significant differences between the methods across all evaluated metrics. These findings reinforce the importance of considering both data characteristics and graph construction strategies when choosing semi-supervised learning techniques.