On the interplay of machine learning and complex networks

• Ano de defesa: 2020
• Autor(a) principal: Mendonça, Matheus Ribeiro Furtado de
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Laboratório Nacional de Computação Científica; Coordenação de Pós-Graduação e Aperfeiçoamento (COPGA); Brasil; LNCC; Programa de Pós-Graduação em Modelagem Computacional
• Palavras-chave em Português: Aprendizado por computador; Teoria dos grafos; Machine Learning-based models; Reinforcement learning (Machine learning); CNPQ::CIENCIAS EXATAS E DA TERRA::MATEMATICA::ALGEBRA
• Link de acesso: https://tede.lncc.br/handle/tede/332

Network Science (NS) and Machine Learning (ML) are two knowledge areas that have experienced exponential growth in recent years, both in academia and in the industry. Both areas deal with broad problems that can be found in different domains, making them a valuable tool for solving such problems. A particular case is how these two research areas interact with each other and what benefits this interaction presents. This thesis thus investigates the interplay between these two areas, in particular, by exploring how NS can be used both as a tool for improving ML methods and also as an application of ML methods. Specifically, we deal with two problems: (i) how to improve Reinforcement Learning (RL) methods using NS; and (ii) how to tackle NS problems using Graph Embedding (GE) and RL. For the first part of our work, we build a new graph-based skill acquisition method, a sub-area of reinforcement learning focused on methods that identify sub-goals within RL problems, and create macro-actions capable of reaching such sub-goals. In this particular context, we propose the Abstract State Transition Graph (ASTG), a more compact representation of the State Transition Graph (STG) used in RL. We then adapt a well-established graph-based skill acquisition method to work with the ASTG instead of the STG. In the second part of our work, we tackle two known NS problems using GE and RL. We first propose the Network Centrality Approximation using Graph Embedding (NCA-GE), a method capable of efficiently approximating node centralities in graphs using GE and supervised learning. The NCA-GE is shown to be better than the state-of-the-art approach, while also being more generic and versatile. We then tackle a different NS problem: efficient information diffusion in Time-Varying Graphs (TVGs). For this task, we propose a method called Spatio-Temporal Influence Maximization (STIM) that uses RL and GE to learn the connectivity and temporal patterns of each node, allowing it to choose the most suitable node to diffuse an information at each time step in a TVG so that this information reaches a significant portion of the nodes within a limited timeframe. The proposed STIM method represents a substantially different approach from any other previous work in the literature, and it establishes a new framework for efficient information diffusion in TVGs that can be adapted for different scenarios.