Aprendizado de máquina para previsão de demanda em e-commerce usando análise de sentimento

Abstract

Data science combines areas such as computing, statistics, and mathematics to process, analyze, and model information, enabling effective predictions to support strategic decision-making. This work proposes the integration of sentiment analysis, a Natural Language Processing (NLP) technique, with machine learning models for demand forecasting in e-commerce. Sentiment analysis was applied to extract polarity (positive, negative, or neutral) from customer reviews, transforming unstructured textual data into structured variables suitable for predictive models. The dataset used was the Brazilian E-Commerce Public Dataset by Olist, which underwent data cleaning, normalization, categorical variable encoding, and class balancing using the SMOTE method. For the modeling stage, three tree-based supervised algorithms were applied: XGBoost, Random Forest, and Extra Trees. The experiments showed that Extra Trees achieved the best performance, with an accuracy of approximately 0.72, followed by Random Forest and, lastly, XGBoost. The results highlighted the common difficulty of all models in classifying neutral sentiments but confirmed the feasibility of integrating sentiment-derived variables into demand forecasting tasks. As a contribution, this study demonstrates that subjective customer feedback can enrich predictive models, enhancing the understanding of consumer behavior and supporting business decision-making.