12 Mar 2021
09:00 Doctoral defense Fully distance
Theme
Learning in non-Euclidean domains: from Graphs to Generative Modeling
Student
Samuel Gomes Fadel
Advisor / Teacher
Ricardo da Silva Torres
Brief summary
This thesis deals with machine learning problems in which data need a non-Euclidean representation, such as graphs, where graph neural networks (GNNs) have been widely used. Our contributions follow three main directions in which we: introduce a graph-based view of problems that are not graph-centered, expand graph problems to a temporal scenario, and take advantage of principles inspired by graph problems, eg Riemann's varieties of constant curvature like hyperspheres, to apply them in new contexts. More specifically, we have introduced an approach to the task of retrieving content with multimodal representations, showing how GNNs can be used as a promising solution to take advantage of information that is not explicitly organized in graphs. Then, we face the recommendation problem, with new representations for temporal changes in graph edges and a model based on GNN, showing that using temporal information leads to better results compared to existing approaches. Furthermore, we use normalizing flows to build movement models that use contextual information to characterize movements in football precisely. We then show how a GNN to be used to, contrary to state-of-the-art approaches, use information from other players, producing even more realistic models. Finally, we investigate problems with interpolations in normalizing flows in their standard scenario, which we attack through a hyperspherical representation, resulting in better quality interpolations. Overall, in our experiments, we get superior performance to alternative approaches. These show how graph-based representations provide a useful means of encoding information, whether this information is extra or expressing how interactions involving thousands of entities develop over time.
Examination Board
Headlines:
Ricardo da Silva Torres | IC / UNICAMP |
Agma Juci Machado Traina | ICMC / USP |
Moacir Antonelli Ponti | ICMC / USP |
Sandra Eliza Fontes de Avila | IC / UNICAMP |
Hélio Pedrini | IC / UNICAMP |
Substitutes:
Nelson Luis Saldanha da Fonseca | IC / UNICAMP |
Jacques Wainer | IC / UNICAMP |
João Paulo Papa | FC / UNESP |