One-Class Learning for Graph Neural Networks

Real-world problems often require identifying a single class of interest (e.g., normal patterns, relevant profiles, confirmed causal relations), while the non-interest space is broad, unknown, heterogeneous, and costly to label. This setting motivates One-Class Learning (OCL), which learns from interest examples and decides whether new instances belong to the target concept. Many of these problems are naturally modeled as graphs, since graphs explicitly represent relationships among entities. In parallel, Graph Neural Networks (GNNs) have become the main approach for learning graph representations by exploiting relational dependencies through neighborhood aggregation, and they are particularly relevant for heterogeneous data. Despite these advances, integrating OCL with GNNs still faces practical limitations: a significant portion of the literature relies on adapting multi-class datasets to one-class settings; unsupervised approaches often yield weak performance; two-step pipelines learn representations that are not customized with the OCL objective; hypersphere-based losses alone may lead to undesirable local minima; and interpretability is commonly handled via post-hoc projections. Moreover, within OCL for GNNs, important problems remain underexplored, such as edge classification and graph streams. To address these gaps, this thesis proposes end-to-end methods that align GNN representation learning with the OCL objective. The core idea is to combine the reconstruction loss of Graph Autoencoders with an OCL sphere loss, explicitly controlling the balance between terms and learning low-dimensional (2D/3D) embeddings to provide interpretability during training. The contributions are organized into four main methods: OLGA for node classification on homogeneous graphs; OLGASTREAM for data streams on homogeneous graphs; ECHOLGA, which explores edge classification for causal discovery via heterogeneous hypergraph modeling; and FOLGA, which explores heterogeneous graphs by leveraging rewiring and graph fusion. Experimental results show that the proposed methods achieve competitive, and often superior, performance compared to representative baselines and state-of-the-art approaches, while also offering interpretability of the learning process, thus advancing the field of One-Class Learning for Graph Neural Networks.