SEPARAÇÃO AUTOMÁTICA DE ATRIBUTOS PARA MÉTODOS DE APRENDIZADO MULTI-VISÃO

Abstract

Multi-view learning is a ``hot'' tendency in machine learning that has produced top-notch results in several applications areas. One of them is automated quality assessment of content created collaboratively on the Web, better exemplified by `Wikis'. Wikis are one of the most common information repositories, to which users resort when they have some information need. Given their free and collaborative nature, such repositories need to control content quality, in order to avoid containing wrong or incomplete information. The state-of-the-art solution for this problem relies on multi-view learning, where quality is considered a multifaceted concept that can be learned from human quality assessments. To this effect, features describing quality have to be devised and grouped into views based on criteria such as text structure, readability, style, user edit history, etc. The task of determining the views requires the assistance of an expert, which is hard to do in scenarios where views are overlapping or hard to interpret by humans. In addition, human engineered views may not be the most adequate for automatically solving the quality measurement problem. In this work, we propose an automatic view generator, to address the problem of generating views for MultiView learning, specially for the problem of automated quality assessment. We evaluate this approach on three popular Wiki datasets. In our experiments, our solution outperformed a version that exploits only the original features, with gains of up to $20$\% in terms of accuracy of the quality assessment. Our method was also able to automatically produce views that are competitive or even better than those manually created, for the task of quality assessment, without any human intervention.