EVENTO
EMBEDDING MULTIDIMENSIONAL NETWORKS
Tipo de evento: Exame de Qualificação
THERE ARE SEVERAL REAL-WORLD COMPLEX SYSTEMS THAT TAKE THE FORM OF NETWORKS, INCLUDING SOCIAL NETWORKS, URBAN TRANSPORTATION SYSTEMS, BIOLOGICAL NETWORKS, AND LANGUAGE NETWORKS. MODELING THESE DOMAINS AS GRAPHS ALLOWS AN UNDERSTANDING OF THE STRUCTURAL INFORMATION ABOUT THESE SYSTEMS BY REPRESENTING INTERACTIONS BETWEEN NODES TO CAPTURE TOPOLOGY PROPERTIES, SUCH AS DEGREE, SHORTEST PATHS, AND CENTRALITIES. HOWEVER, REAL COMPLEX NETWORKS OFTEN INCORPORATE PROPERTIES NOT COVERED BY TRADITIONAL GRAPH THEORY, SUCH AS TOPOLOGICAL EVOLUTION, LAYER STRUCTURE, AND MULTI-SCALED PATTERNS, THEREFORE GRAPH GENERALIZATIONS HAVE BEEN PROPOSED OVER THE LAST 10 YEARS IN ORDER TO BETTER REPRESENT THESE MULTIDIMENSIONAL NETWORKS. FROM TIME-VARYING GRAPH MODELS, EXTENDING STATIC GRAPHS TO HANDLE DYNAMIC NETWORKS, TO COMPLEX SYSTEMS COMPOSED OF A SET OF INTERDEPENDENT NETWORKS COLLABORATING FOR THE WHOLE PICTURE OF THE SYSTEM, THUS FORMING MULTI-LAYER NETWORKS, SEVERAL GRAPH GENERALIZATIONS WERE PROPOSED. DEFINING AND MEASURING TOPOLOGICAL PROPERTIES ALONG WITH TIME EVOLUTION AND LAYER PATTERNS IN HIGHER ORDER NETWORKS ARE MATHEMATICALLY AND COMPUTATIONALLY CHALLENGING TASKS. THEREFORE, NETWORK EMBEDDING ARISES AS A MACHINE LEARNING TASK, AIMING TO LEARN REPRESENTATIONS FROM DATA BY MAPPING A GRAPH INTO A LOW-DIMENSIONAL VECTOR SPACE IN WHICH SEMANTIC, RELATIONAL, AND STRUCTURAL INFORMATION CAN BE CAPTURED. APPLYING THESE REPRESENTATIONS AS FEATURES TO A MACHINE LEARNING MODEL ENABLES PERFORMING TASKS SUCH AS NODE CLASSIFICATION, RANKING AND CLUSTERING, LINK PREDICTION, AND GRAPH CLASSIFICATION. HOWEVER, EXISTING WORKS ON NETWORK EMBEDDING PRIMARILY FOCUS ON STATIC GRAPHS, AND ALTHOUGH THERE IS A RECENT RESEARCH EFFORT IN THE AREA OF EMBEDDING DYNAMIC GRAPHS, FULLY ESTABLISHED AND CONSENSUAL DEFINITIONS ARE NEEDED SO FAR. EVEN FURTHER, WHEN REFERRING TO MULTIDIMENSIONAL NETWORKS, WITH LAYERS AND MULTI-SCALE STRUCTURES, DEVELOPING TECHNIQUES TO MAP THEM IN VECTOR SPACES BECOMES EVEN MORE CHALLENGING. IN VIEW OF THIS PROBLEM, THE PRESENT PROPOSAL AIMS AT THE THEORETICAL AND ALGORITHMIC DEVELOPMENT OF EMBEDDING-BASED SOLUTIONS APPLIED TO HIGHER ORDER DIMENSIONAL NETWORKS. GIVEN THE DIFFERENCES AMONG MULTIDIMENSIONAL GRAPH MODELS (I.E. TENSOR REPRESENTATION GRAPHS, STREAM GRAPHS AND MULTIASPECT GRAPHS), AFTER A BIBLIOGRAPHIC SURVEY OF EACH ONE, A DISCUSSION WILL BE MADE ON HOW TO ADAPT THE STATE-OF-THE-ART METHODS AND EXPAND THEM TO MULTIDIMENSIONAL NETWORKS. SEVERAL EMBEDDING PARADIGMS, INCLUDING MATRIX FACTORIZATION, DEEP LEARNING AND RANDOM WALKS, WILL BE EVALUATED IN HIGHER ORDER NETWORK SCENARIOS, IN ORDER TO ENABLE FURTHER APPLICATIONS SUCH AS EVENT TIME PREDICTION, TIME CENTRALITY, LAYER CENTRALITY AND MULTI-SCALE INFORMATION DIFFUSION. THEREFORE, THE PRESENT WORK WILL PROVIDE AN EXPANSION OF THE SCIENTIFIC COMMUNITY'S CURRENT KNOWLEDGE ABOUT REPRESENTATION LEARNING OVER NETWORKS BY ALLOWING NEW EMBEDDING TECHNIQUES TO BE DEVELOPED, NEW APPLICATION SCENARIOS TO BE EVALUATED AND, MOREOVER, THE DEVELOPMENT OF AN EFFICIENT AND ACCURATE METHODOLOGY FOR NETWORK ANALYTICS AND INFERENCE PROBLEMS. Para assistir essa defesa acesse o link: https://us02web.zoom.us/j/87035962936?pwd=dVB0SkEydk11c2ZDWXIrc0pENER1UT09
Local: LNCC - Laboratório Nacional de Computação Ciêntifica
Endereço: Getúlio Vargas Av., 333, Quitandinha Petrópolis - Rio de Janeiro CEP 25651-075 - Brasil
Telefone: (24) 2233.6004
Data Início: 09/07/2020 Data Fim: 09/07/2020
Aluno: Claudio Daniel Tenório de Barros - - LNCC
Orientador: Artur Ziviani - Laboratório Nacional de Computação Científica - LNCC
Participante Banca Examinadora: Alex Borges Vieira - Universidade Federal de Juiz de Fora - UFJF Fabio André Machado Porto - Laboratório Nacional de Computação Científica - LNCC Pablo Javier Blanco - Laboratório Nacional de Computação Científica - LNCC
Suplente Banca Examinadora: Antônio Tadeu Azevedo Gomes - Laboratório Nacional de Computação Científica - LNCC
