Comparando algoritimos multirrótulos aplicados em mineração de sentimentos

Abstract

The advancement of the internet and the increase in the number of users in recent years have led to a considerable rise in the amount of available data. This data has become extremely valuable for large global companies, which use it as the basis for Machine Learning algorithms. These algorithms are employed in analyzing their target audience's opinions regarding the brand and products, as well as in various activities on their platforms, such as search engine operations, product recommendations, spam filtering, and targeted advertising, among others. The main objective of this study is to compare different multi-label classification algorithms that use problem transformation techniques, including Binary Relevance, Classifier Chains, Random k-Labelsets, and Label Powerset. These algorithms are applied to three sentiment datasets: the first is a dataset with tweets containing 6179 entries in English and 11 unbalanced labels, the second with 5131 comments about depression on a subreddit and 8 balanced labels, and the third with 30,000 comments from various subtopics on Reddit and 27 unbalanced labels. The purpose is to apply metrics such as accuracy, precision, recall, f1-score, and precision at K for each algorithm, determining their performance under varied conditions, implementing these algorithms using the Python programming language and the Scikit-Multilearn library. When comparing the results, it is noted that the Binary Relevance algorithm performs superiorly compared to the others when trained on a dataset with balanced labels, standing out mainly in precision with an approximate result of 75% and in precision at K with 69%. On the other hand, the Classifier Chains algorithm shows consistent performance across all datasets, notably with its precision of 76%. Despite a significant drop in performance on an unbalanced dataset, the Random k-Labelsets and Label Powerset algorithms demonstrate better performance, particularly standing out in accuracy with results close to 31% for both and in recall with around 22% for both again.