Relevant data representation by a Kernel-based framework

Nowadays, the analysis of a large amount of data has emerged as an issue of great interest taking increasing place in the scientific community, especially in automation, signal processing, pattern recognition, and machine learning. In this sense, the identification, description, classification, visualization, and clustering of events or patterns are important problems for engineering developments and scientific issues, such as biology, medicine, economy, artificial vision, artificial intelligence, and industrial production. Nonetheless, it is difficult to interpret the available information due to its complexity and a large amount of obtained features. In addition, the analysis of the input data requires the development of methodologies that allow to reveal the relevant behaviors of the studied process, particularly, when such signals contain hidden structures varying over a given domain, e.g., space and/or time. When the analyzed signal contains such kind of properties, directly applying signal processing and machine learning procedures without considering a suitable model that deals with both the statistical distribution and the data structure, can lead in unstable performance results. Regarding this, kernel functions appear as an alternative approach to address the aforementioned issues by providing flexible mathematical tools that allow enhancing data representation for supporting signal processing and machine learning systems. Moreover, kernelbased methods are powerful tools for developing better-performing solutions by adapting the kernel to a given problem, instead of learning data relationships from explicit raw vector representations. However, building suitable kernels requires some user prior knowledge about input data, which is not available in most of the practical cases. Furthermore, using the definitions of traditional kernel methods directly, possess a challenging estimation problem that often leads to strong simplifications that restrict the kind of representation that we can use on the data. In this study, we propose a data representation framework based on kernel methods to learn automatically relevant sample relationships in learning systems. Namely, the proposed framework is divided into five kernel-based approaches, which aim to compute relevant data representations by adapting them according to both the imposed sample relationships constraints and the learning scenario (unsupervised or supervised task). First, we develop a kernel-based representation approach that allows revealing the main input sample relations by including relevant data structures using graph-based sparse constraints. Thus, salient data structures are highlighted aiming to favor further unsupervised clustering stages. This approach can be viewed as a graph pruning strategy within a spectral clustering framework which allows enhancing both the local and global data consistencies for a given input similarity matrix. Second, we introduce a kernel-based representation methodology that captures meaningful data relations in terms of their statistical distribution. Thus, an information theoretic learning (ITL) based penalty function is introduced to estimate a kernel-based similarity that maximizes the whole information potential variability. So, we seek for a reproducing kernel Hilbert space (RKHS) that spans the widest information force magnitudes among data points to support further clustering stages. Third, an entropy-like functional on positive definite matrices based on Renyi’s definition is adapted to develop a kernel-based representation approach which considers the statistical distribution and the salient data structures. Thereby, relevant input patterns are highlighted in unsupervised learning tasks. Particularly, the introduced approach is tested as a tool to encode relevant local and global input data relationships in dimensional reduction applications. Fourth, a supervised kernel-based representation is introduced via a metric learning procedure in RKHS that takes advantage of the user-prior knowledge, when available, regarding the studied learning task. Such an approach incorporates the proposed ITL-based kernel functional estimation strategy to adapt automatically the relevant representation using both the supervised information and the input data statistical distribution. As a result, relevant sample dependencies are highlighted by weighting the input features that mostly encode the supervised learning task. Finally, a new generalized kernel-based measure is proposed by taking advantage of different RKHSs. In this way, relevant dependencies are highlighted automatically by considering the input data domain-varying behavior and the user-prior knowledge (supervised information) when available. The proposed measure is an extension of the well-known crosscorrentropy function based on Hilbert space embeddings. Throughout the study, the proposed kernel-based framework is applied to biosignal and image data as an alternative to support aided diagnosis systems and image-based object analysis. Indeed, the introduced kernel-based framework improve, in most of the cases, unsupervised and supervised learning performances, aiding researchers in their quest to process and to favor the understanding of complex data

• Autores: Álvarez Meza, Andrés Marino
• Tipo de recurso: Doctoral thesis
• Fecha de publicación: 2015
• Institución: Universidad Nacional de Colombia
• Repositorio: Universidad Nacional de Colombia
• Idioma: spa
• OAI Identifier: oai:repositorio.unal.edu.co:unal/76341
• Acceso en línea: https://repositorio.unal.edu.co/handle/unal/76341; http://bdigital.unal.edu.co/72629/
• Palabra clave: Signal processing; Machine learning; Relevant representation; Kernel methods; Information theoretic learning; Automatics; Procesamiento de señales; Aprendizaje de máquina; Representación relevante; Métodos núcleo; Aprendizaje por teoría de información; Automatización
• Rights: openAccess
• License: Atribución-NoComercial 4.0 Internacional