Redução de dimensionalidade em predições biométricas
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidade Federal de Viçosa
Abstract
A computação aplicada tem um papel crucial no Melhoramento Genético seja no se- quenciamento e análise de genoma, na identificação de marcadores moleculares ou no aprimoramento de técnicas de inteligência computacional. Um dos domínios do Melhoramento Genético que tem feito extenso uso de dados reais e simulação é a Bi- ometria. Um dos focos da Biometria é a seleção de modelos eficientes para o processo de Seleção Genômica Ampla (GWS) utilizando marcadores moleculares (SNPs) e esse procedimento apresenta vários desafios. Esse estudo buscou fornecer respostas a dois (02) desses desafios: a) comparar eficiência de diferentes técnicas de Predição em Bio- metria para identificação de marcadores moleculares (SNPs) relevantes para controle genético de características simuladas; b) estabelecer procedimentos de Redução de Di- mensionalidade (ou Feature Selection) de conjunto de marcadores moleculares (SNPs) para fins de Predição de características complexas simuladas, considerando intera- ção gênica, dominância e herdabilidade diferenciada. Para elucidar tais problemas foram gerados dados por simulação (via programa Genes) do tipo marcadores mo- leculares (Single Nucleotide Polymorphisms - SNPs) e esse conjunto de dados reflete os principais problemas encontrados nesta linha de investigação (alta dimensionalidade, não-linearidade e multicolinearidade). A comparação da eficiência foi feita através da avaliação da acurácia seletiva de nove (09) modelos de Predição que seguem di- ferentes paradigmas no contexto da Biometria. Como principais resultados, pode-se destacar o aumento da eficiência preditiva à medida que o ruído dos dados diminui, a superioridade do paradigma da árvore (para baixos níveis de ruído, BOO) e a efi- ciência do paradigma da rede neural (para altos níveis de ruído, RBF). O segundo desafio teve como ponto de partida o fato de que Redução de Dimensionalidade (RD) tem se tornado uma ferramenta fundamental para melhorar a eficiência de modelos de Predição e mais especificamente ainda a classe de modelos do tipo Feature Selection (FS). Visando contribuir com esse processo, este estudo fez uso três (03) técnicas de Feature Selection (Bagging, Sonda e Stepwiswe) aplicadas a um modelo do tipo (Random Forest - RF). Os resultados obtidos sinalizam melhorias na acurácia seletiva (R2 ) entre 10% e 28% e uma melhor adequação de dois (02) modelos (Bagging e Stepwiswe) em detrimento a um modelo (Sonda), em captar de maneira mais adequada as situações que envolvem o controle gênico de uma característica. Por fim,a medida que o ruído diminui, a acurácia seletiva aumenta (isso vale para modelos com e sem RD) e as taxas de crescimento da acurácia seletiva se tornam decrescentes. Palavras-chave: Modelos de Predição. Redução de Dimensionalidade. Feature Selec- tion. SNPs. Biometria.
Applied computing plays a crucial role in Genetic Improvement, whether in genome sequencing and analysis, the identification of molecular markers or the improvement of computational intelligence techniques. One of the areas of breeding that has made extensive use of real data and simulation is biometrics. One of the focuses of Biome- trics is the selection of efficient models for the Genome-Wide Selection (GWS) process using molecular markers (SNPs) and this procedure presents several challenges. This study sought to provide answers to two (02) of these challenges: a) to compare the effi- ciency of different Biometrics Prediction techniques for identifying relevant molecular markers (SNPs) for genetic control of simulated traits; b) to establish procedures for Dimensionality Reduction (or Feature Selection) of a set of molecular markers (SNPs) for the purposes of Prediction of complex simulated traits, considering gene interac- tion, dominance and differentiated heritability. In order to elucidate these problems, simulation data was generated (via the Genes program) for molecular markers (Sin- gle Nucleotide Polymorphisms - SNPs) and this data set reflects the main problems encountered in this line of research (high dimensionality, non-linearity and multicol- linearity). Efficiency was compared by evaluating the selective accuracy of nine (09) prediction models that follow different paradigms in the context of Biometrics. The main results were the increase in predictive efficiency as data noise decreased, the superiority of the tree paradigm (for low levels of noise, BOO) and the efficiency of the neural network paradigm (for high levels of noise, RBF). The second challenge was based on the fact that Dimensionality Reduction (DR) has become a fundamental tool for improving the efficiency of prediction models, and more specifically the class of Feature Selection (FS) models. In order to contribute to this process, this study used three (03) Feature Selection techniques (Bagging, Sonda and Stepwiswe) applied to a Random Forest (RF) model. The results obtained indicate improvements in selec- tive accuracy (R2 ) between 10% and 28% and a better suitability of two (02) models (Bagging and Stepwiswe) to the detriment of one model (Sonda), in more adequately capturing situations involving the genetic control of a trait. Finally, as noise decre- ases, selective accuracy increases (this is true for models with and without RD) and selective accuracy growth rates become decreasing.Keywords: Prediction Models. Dimensionality Reduction. Feature Selection. SNPs. Biometric.
Applied computing plays a crucial role in Genetic Improvement, whether in genome sequencing and analysis, the identification of molecular markers or the improvement of computational intelligence techniques. One of the areas of breeding that has made extensive use of real data and simulation is biometrics. One of the focuses of Biome- trics is the selection of efficient models for the Genome-Wide Selection (GWS) process using molecular markers (SNPs) and this procedure presents several challenges. This study sought to provide answers to two (02) of these challenges: a) to compare the effi- ciency of different Biometrics Prediction techniques for identifying relevant molecular markers (SNPs) for genetic control of simulated traits; b) to establish procedures for Dimensionality Reduction (or Feature Selection) of a set of molecular markers (SNPs) for the purposes of Prediction of complex simulated traits, considering gene interac- tion, dominance and differentiated heritability. In order to elucidate these problems, simulation data was generated (via the Genes program) for molecular markers (Sin- gle Nucleotide Polymorphisms - SNPs) and this data set reflects the main problems encountered in this line of research (high dimensionality, non-linearity and multicol- linearity). Efficiency was compared by evaluating the selective accuracy of nine (09) prediction models that follow different paradigms in the context of Biometrics. The main results were the increase in predictive efficiency as data noise decreased, the superiority of the tree paradigm (for low levels of noise, BOO) and the efficiency of the neural network paradigm (for high levels of noise, RBF). The second challenge was based on the fact that Dimensionality Reduction (DR) has become a fundamental tool for improving the efficiency of prediction models, and more specifically the class of Feature Selection (FS) models. In order to contribute to this process, this study used three (03) Feature Selection techniques (Bagging, Sonda and Stepwiswe) applied to a Random Forest (RF) model. The results obtained indicate improvements in selec- tive accuracy (R2 ) between 10% and 28% and a better suitability of two (02) models (Bagging and Stepwiswe) to the detriment of one model (Sonda), in more adequately capturing situations involving the genetic control of a trait. Finally, as noise decre- ases, selective accuracy increases (this is true for models with and without RD) and selective accuracy growth rates become decreasing.Keywords: Prediction Models. Dimensionality Reduction. Feature Selection. SNPs. Biometric.
Description
Citation
GUIMARÃES, Patrick Wöhrle. Redução de dimensionalidade em predições biométricas. 2024. 54 f. Dissertação (Mestrado em Ciência da Computação) - Universidade Federal de Viçosa, Viçosa. 2024.
