Classificação de itens em compras públicas com uso de algoritmos de aprendizagem de maquina

Abstract

Text classification is an area of application of machine learning that has gained increasing importance, especially after the spread of computers, with the use of digital documents, and the internet, especially with the growth of social networks. Despite the availability of data, given the lack of standardization in the description of items, municipalities are unable to identify their consumption of certain items, or categories of items. As a result, they are unable to plan adequately, making the public budget less efficient. The objective of the work is to develop short text classifiers to classify items tendered by municipalities in Paraná. In summary, the results obtained in this study highlight the effectiveness of KNN in classifying bid items based on their descriptions. The KNN TF-IDF Model presented the best performance, with an accuracy of 85.21% on the test basis. KNN’s superior performance highlights its ability to capture close relationships in feature spaces, proving it to be a promising choice for similar tasks in municipal bidding contexts. Furthermore, it is worth highlighting that the SVM and NB models also demonstrated promising results, indicating the feasibility of alternative approaches in the classification of tendered items. It is noteworthy that the models that used the LSA and LDA feature extraction methods did not present inferior results to the models that used BoW and TF-IDF. Thus, the results indicate that using LSA and LDA for feature extraction is possibly not an efficient approach for the task proposed in this work.