Abordagem bioinspirada híbrida de seleção de atributos para classificação de sentimentos em mídias sociais
Carregando...
Arquivos
Data
2022-04-29
Autores
Silva, Luis Marcello Moraes [UNESP]
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
A análise de sentimentos em mídias sociais consiste em extrair informações de usuários presentes nos comentários destas redes sociais. Este tema tem sido amplamente estudado nos últimos anos, pois, por exemplo, pode auxiliar o processo de tomada de decisão de empresas e até identificar intenções e opiniões sobre candidatos em eleições. No entanto, devido ao ambiente Big Data no qual estes dados estão inseridos, sua análise tradicional pode ficar comprometida por conta do elevado número de atributos somados a outros fatores. Isto acaba por implicar em uma análise com alto custo computacional e com baixa qualidade de resultados, além do fato de que tal análise é inviável manualmente, pois excede a capacidade humana de entendimento. Pesquisas recentes têm focado em como analisar os sentimentos de usuários com técnicas de aprendizado de máquina somadas às técnicas inspiradas pela natureza, e assim, possibilitar o estudo de opiniões de usuários sobre um determinado tópico. Com o intuito de se analisar tais dados de modo mais preciso, uma seleção de atributos por meio destas abordagens, somado a análises léxicas, tornou-se uma alternativa atrativa para contornar este desafio e viabilizar seu processamento. Este trabalho tem como objetivo a apresentação de uma abordagem híbrida bioinspirada, cuja contribuição científica é a melhoria de um modelo preditivo de classificação de sentimentos multi-idiomas que considera diferentes contextos dos dados. Por meio dos resultados, é possível verificar que o modelo obteve melhorias de acurácia entre 10% e 17%, enquanto que o método de seleção utilizou cerca de 45% dos atributos em relação à análise tradicional.
The social media sentiment analysis consists on extracting information from users in comments made in their social network. Such topic has been the focus of many study works in the last few years. It can assist the decision-making process of companies, aid teaching methods and even identify and boost intentions and opinions about candidates in elections. However, due to the Big Data environment in which these data are inserted, the traditional analysis can be compromised because of the high dimensionality added to other factors. The implication on the analysis is resulted by high computational cost and low quality of results. Besides that such analysis is impracticable manually as it exceeds the human capacity of understanding. Up to date research has given a focus on how to analyze feelings of users with machine learning and techniques inspired by nature, allowing the study of users’ opinions. In order to analyze such data effectively, a feature selection through these approaches is proposed. Machine learning added to lexical analysis has become an attractive alternative to overcome this challenge and facilitate its processing. This paper aims to present a hybrid bioinspired approach to realize feature selection and improve sentiment classification quality. The scientific contribution is the improvement of a classification model considering preprocessing of the data with different languages and contexts. The results prove that the developed method enriches the predictive model by improving the accuracy by 10% to 17%. This method selected 45% of the attributes in average compared to traditional analysis.
The social media sentiment analysis consists on extracting information from users in comments made in their social network. Such topic has been the focus of many study works in the last few years. It can assist the decision-making process of companies, aid teaching methods and even identify and boost intentions and opinions about candidates in elections. However, due to the Big Data environment in which these data are inserted, the traditional analysis can be compromised because of the high dimensionality added to other factors. The implication on the analysis is resulted by high computational cost and low quality of results. Besides that such analysis is impracticable manually as it exceeds the human capacity of understanding. Up to date research has given a focus on how to analyze feelings of users with machine learning and techniques inspired by nature, allowing the study of users’ opinions. In order to analyze such data effectively, a feature selection through these approaches is proposed. Machine learning added to lexical analysis has become an attractive alternative to overcome this challenge and facilitate its processing. This paper aims to present a hybrid bioinspired approach to realize feature selection and improve sentiment classification quality. The scientific contribution is the improvement of a classification model considering preprocessing of the data with different languages and contexts. The results prove that the developed method enriches the predictive model by improving the accuracy by 10% to 17%. This method selected 45% of the attributes in average compared to traditional analysis.
Descrição
Palavras-chave
Ciência da computação, Inteligência coletiva, Algoritmos genéticos, Redes sociais, Sistemas especialistas (Computação), Computer science, Swarm intelligence, Genetic algorithms, Social networks, Knowledge acquisition (Expert systems)