Defesa – Tese de Silvio Fernando Angonese
Aluno(a): Silvio Fernando Angonese
Orientador(a): Renata de Matos Galante
Título: AGHE – Approach for Generating Enhanced Heterogeneous Node Embeddings in Heterogeneous Graphs
Linha de Pesquisa: Mineração, Integração e Análise de Dados
Data: 14/05/2026
Hora: 09:00
Local: Esta banca ocorrerá de forma remota. Acesso público disponibilizado pelo link https://meet.google.com/aja-uhaz-oir.
Banca Examinadora:
-Anderson Rocha Tavares (UFRGS)
-Damires Souza (IFPB)
-Carina F. Dorneles (UFSC)
Presidente da Banca: Renata de Matos Galante
Resumo: Heterogeneous graphs provide a powerful framework for modeling complex relationships across multiple modalities, such as text, images, and structured subgraphs. A key research question addressed in this thesis is whether it is possible to construct a heterogeneous graph enriched with multiple node representations, integrating multimodal data into single and embedding compositions, in a coherent and structured manner that can effectively support node classification tasks. To tackle this challenge, we propose AGHE (Approach for Generating Enhanced Heterogeneous Node Embeddings), a unified approach for generating and composing heterogeneous embeddings. AGHE integrates local node features, aggregated neighbor information, and metapath-based semantics into enriched vector representations, enabling more expressive and informative node classification models. Experimental evaluations conducted across three complementary settings demonstrate the effectiveness of AGHE in constructing semantically enriched heterogeneous graphs and generating embeddings fully compatible with supervised classification models. The results show that embedding compositions consistently outperform single embeddings, leading to improved predictive performance and stronger separability. Moreover, controlled structural ablation analyses indicate that composition-based embeddings maintain greater stability under progressive relational simplification, evidencing enhanced robustness and generalization across heterogeneous scenarios. These findings confirm both the validity of the AGHE approach for heterogeneous embedding generation and the importance of composition strategies as a principled foundation for advancing representation learning in heterogeneous graphs.
Palavras-Chave: Heterogeneous Embedding. Multi-modal Features. Heterogeneous Graph. Graph Embedding. Composition Embeddings. Graph Neural Networks.