A Thorough exploitation of distance-based meta-features for Automated text classification

• Autor(a) principal: Sergio Daniel Carvalho Canuto
• Data de Publicação: 2019
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Repositório Institucional da UFMG
• Texto Completo: http://hdl.handle.net/1843/34071
• topic: Supervised classification; Text classification; Meta-features; Machine learning; Computação – Teses; Aprendizado supervisionado.; Meta características; Aprendizado de máquina

Automated Text Classification (ATC) has become substantially important for a variety of tasks, such as categorizing news, organizing digital libraries, building web directories, analyzing sentiment of user-generated content and detecting spam, to name a few. Given a set of training documents classified into one or more predefined categories, the task of ATC is to utomatically learn how to classify new (unclassified) documents, using a combination of features of these documents that associates them with categories. Due to the fact that the ATC problem occurs in a number of different applications, diverse machine learning algorithms have been proposed to deal with ATC. Although the classification algorithm itself plays an important role in ATC, the features that represent documents may be equally important to determine effectiveness. In particular, representing documents in a feature space is a prerequisite work for ATC, since these classification algorithms are designed to discover discriminative patterns on these features. In this sense, a relevant challenge relies on efficiently manipulating the feature space to address ATC from a data engineering viewpoint. In this context, we address the problem of automatically learning to classify texts by exploiting information derived from meta-features, i.e., features engineered from the original (bag-of-words) representation. Particularly, the exploited meta-features rely on distance measures to summarize complex relationships between documents and present discriminative information for classification. We here not only propose new meta-features that provide discriminative evidence for classification, but also new mechanisms to analyze and select meta-features using multi-objective strategies. These strategies are capable of reducing the number of meta-features while maximizing the classification effectiveness, when considering the adequacy of the selected meta-features to a particular dataset or classification method. Moreover, we provide additional contributions to improve the efficiency and effectiveness of meta-features. Particularly, we propose: (i) the use of commodity GPUs to reduce the computational time to generate meta-features; (ii) the use of supervised learning to enrich distance relationships with labeled information; and (iii) the design of new specific meta-features for the sentiment analysis context. Our experimental results on five traditional benchmarks for topic classification show that with the appropriate selection techniques, our distance-based meta-features can achieve remarkable classification results considering the results of original feature space and other recently proposed distance-based meta-features. We further explain our results with the identification and discussion about meta-features that, when combined, provide core information to classify documents. Our improvements on core meta-features using labeled information to enrich distance relationships provide additional gains over our best results in topic datasets. We also evaluate meta-features on nineteen sentiment analysis datasets. In this context, our proposals for sentiment classification produced remarkable results considering the effectiveness of previous meta-features that do not take sentiment analysis idiosyncrasies into account.