Aprendizado positivo e não rotulado guiado por densidade

Abstract

Positive-Unlabeled Learning is a special case of semi-supervised learning in which only a few examples of the positive class are labeled, while the rest remain unlabeled, comprising both positive and negative instances. The two-step strategy, widely adopted in the literature, consists of extracting a subset of reliable negative examples from the unlabeled data and subsequently training a classifier based on them. This work proposes a modular framework based on density-based hierarchical clustering for the extraction of reliable negatives and for a novel transductive semi-supervised classification strategy. The approach is founded on the exploration of the data's topological structure and on density-based probability estimates. To extend the method to data represented as graphs, the strategy is combined with latent representation learning techniques. An adaptation focused on topology preservation is proposed for scenarios where connected nodes tend to be similar (homophily), alongside a novel spatially guided metric learning architecture, developed in this work for networks where connections between distinct classes are predominant (heterophily). The experimental validation encompassed several datasets, contrasting the method with recent state-of-the-art approaches. The findings corroborate the premise that density in the latent space is a strong indicator of class separability in the absence of negative labels, ensuring the proposed approach achieves competitive, and sometimes superior, performance in both negative extraction and classification across tabular data and homophilic graphs. The scenario proves more challenging in strongly heterophilic networks, where the developed model preserves a good capacity for negative extraction but encounters limitations regarding the complexity of semi-supervised propagation during the final classification. This dynamics in complex graphs, coupled with the model's sensitivity to structural parameters, outlines the current application limits of the framework and points to natural directions for future work.