Full text
| |
| Author(s): |
Samuel Gomes Fadel
Total Authors: 1
|
| Document type: | Doctoral Thesis |
| Press: | Campinas, SP. |
| Institution: | Universidade Estadual de Campinas (UNICAMP). Instituto de Computação |
| Defense date: | 2021-03-12 |
| Examining board members: |
Ricardo da Silva Torres;
Agma Juci Machado Traina;
Moacir Antonelli Ponti;
Sandra Eliza Fontes de Avila;
Hélio Pedrini
|
| Advisor: | Ricardo da Silva Torres |
| Abstract | |
This thesis deals with machine learning problems where data requires a representation in non-Euclidean domains, such as graphs. Our contributions follow three main directions, in which we: introduce a graph-based view to problems that are not graph-centric, expand graph problems to a temporal setting, and take advantage of principles inspired by graph problems, e.g., constant-curvature Riemannian manifolds, such as hyperspheres, to employ them in novel ways. More specifically, we introduce an approach for retrieving content with multimodal representations, showing how graph neural networks (GNNs) can be used as a promising solution to leverage information not explicitly organized as a graph. Next, we address the recommendation problem, introducing representations for temporal changes in graph edges and a GNN-based model, showing that leveraging temporal information outperforms prior approaches. Furthermore, we use a normalizing flow to build movement models that can use contextual information to precisely characterize movements in soccer. We then show how a GNN can be used to leverage information about other players, producing even more realistic movement models. Lastly, we investigate interpolation issues in normalizing flows, which we address by using a hyperspherical representation, leading to interpolations of higher quality. In our experiments, the methods here proposed obtain better performance over alternative approaches, showing that graph representations provide useful means for encoding information, either about context or about how interactions involving thousands of entities unfold over time. (AU) | |
| FAPESP's process: | 17/24005-2 - Temporal relational reasoning with neural networks |
| Grantee: | Samuel Gomes Fadel |
| Support Opportunities: | Scholarships in Brazil - Doctorate |