Comparison of Machine Learning Models for Sentiment Analysis of Big Turkish Web-Based Data

E-commerce sites have generated large amounts of unstructured data as they allow millions of users to generate product reviews. Thus, although there have been significant improvements in the characteristics of big data, such as speed and volume, developing various analysis techniques to monitor, understand, and extract useful information from this web-based data has become challenging. This study aims to analyze cosmetic products on a Turkish-based e-commerce website with sentiment analysis and to create a new domain-specific Turkish sentiment dictionary model with manual labeling. In the study, a Turkish sentiment dictionary consisting of 65,378 words was created by manually labeling 875,455 product reviews for 24 cosmetic brands sold on the Turkey-based trendyol e-commerce site, and sentiment analysis was performed using this dictionary. The dataset, divided into seven product groups, was analyzed using K-NN, SVM, DT, RF, and LR algorithms to address three classification problems. The algorithms were evaluated with comparative analysis using accuracy, precision, recall, and f-1 score metrics. SVM gave the highest performance result with over 93% accuracy, 92% precision, 93% recall, and a 91% f-1 score in all product groups. The dictionary model created for the cosmetics industry in the study helps businesses and researchers to use their resources more efficiently and save time by performing fast and low-cost analyses on large datasets of product reviews. Moreover, by analyzing customer feedback, brands can offer long-lasting and environmentally friendly products that align with customers’ feelings. Thus, businesses have the opportunity to develop or improve products.