Development of complex network community detection techniques and applications in invariant pattern recognition

Abstract

Machine learning aims at the development of computational methods capable of "learn" with experience. Such techniques are able to, on an automatic form, generate models to organize the existing knowledge or even imitate the behavior of a human expert in certain rank. Another interdisciplinary domain with intense development in the last years is the complex networks. Such networks have emerged recently as a unified topic in complex systems present in several branches of the science. This project has objective to develop machine-learning techniques for data analysis based on complex networks and dynamical systems. Specifically, we will study and develop community detection techniques in complex networks based on particle competition and synchronization of coupled oscillators, which can serve as non-supervised learning techniques (clustering) with the capacity to detect clusters of varied forms and the hierarchical representation. Then, an adaptation of the developed techniques to fit semi-supervised learning will be carried out. In this case, some nodes of the network are labeled, i.e., their classes are previously defined. The model to be developed should be capable of propagating the labels to others vertices of the network. Finally, the semi-supervised learning techniques will be applied to treat problems of invariant pattern recognition, mainly in the case of the presence of nonlinear distortions in the patterns under analysis. The main idea is to recognize a pattern with certain variances through propagation of its label to the corresponding stored pattern via intermediate patterns, all of them represented by nodes of a network. This approach is denoted as relational pattern recognition. Complex networks and dynamical systems are powerful tools for many disciplines of the science, including for machine learning, and there still exists a large space for exploitation. Therefore, we believe that the use of complex networks and dynamical systems can offer a good contribution for invariant pattern recognition. (AU)