Understanding the search space of methods for automatically designing graph neural networks

• Autor(a) principal: Matheus Henrique do Nascimento Nunes
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Institucional da UFMG
• Texto Completo: http://hdl.handle.net/1843/47526; https://orcid.org/0000-0001-5975-7903
• topic: Graph Neural Networks; Automated Machine Learning; Neural Architecture Search; Computação - Teses; Redes neurais ( Computação) - Teses; Aprendizado de máquina - Teses

Graph-structured data has become increasingly available and, due to its ubiquity, an object of study in many areas of research. Due to the absence of the notion of sequence in graphs, Machine Learning (ML) methods have historically struggled to work on this data. Specialized methods for performing ML over graph data have gained a lot of attention from the research community, especially Graph Neural Networks (GNNs), which have been extensively used over real-world data, achieving state-of-the-art results in tasks such as circuit design, movie recommendation, and anomaly detection. Many GNN models have been recently proposed, and choosing the best model for each problem has become a cumbersome and error-prone task. Aiming at mitigating this problem, recent works have proposed strategies for applying Neural Architecture Search (NAS) - a set of methods designed to automatically configure neural networks, very successful on Convolutional Neural Networks, that deal with image data - to GNN models. Automatically configured GNNs have achieved high performance results, surpassing human-crafted ones. However, the NAS for GNNs literature is still in its early stages, and methods that have been successfully applied for NAS in CNNs have yet to be tested on GNNs as well. In this work we have conducted a comprehensive comparative analysis of a proposed Evolutionary Algorithm against a literature Reinforcement Learning and a simple Random Search baseline, considering 7 real-world datasets and two search spaces. We have shown that Random Search is just as effective in finding good performing architectures as other more complex methods. Another interesting finding is that all three search methods converge very early in the search (in about 10% of the budget). We hypothesized that this might have been happening due to the presence of Neutrality (regions in which all solutions have very similar performance values) in the search space. Shifting the focus from the first part of this work, in the second part we have conducted an extensive visual and analytical evaluation of one of the literature's search spaces, using dimensionality reduction and Fitness Landscape Analysis techniques. We have demonstrated that the search space for NAS in GNNs presents high searchability (i.e. it is not difficult for algorithms to explore and find a suitable solution) and neutrality (i.e. there are many regions in the search space in which the performance of the neighboring solutions are relatively equal). We hypothesize that in the future, less expensive methods can be used to perform the optimization in this scenario without loss of generality.