Abstract:

In recent years, there has been a rapid evolution in e-commerce platforms, facilitating global transactions of goods and services over the Internet. Categorizing shopping app reviews based on specific features provides valuable support for well-informed decision-making in a business setting. This includes aspects such as app maintenance, optimization, evolution, and prioritizing user requests, which, in turn, helps in improving user retention and increasing in-app revenue. In this paper, we propose a taxonomy to classify users' reviews from shopping apps into five quality-centered classifications: system quality, information quality, service quality, product quality, and delivery quality. Subsequently, we created an initial dataset containing 552 instances and employed five common multi-label classification models: binary relevance, classifier chain, label powerset, multi-label k nearest neighbors, and multi-label decision tree. The outcomes of this preliminary work are promising, achieving a Hamming loss of 0.177, Macro-F1 score of 0.739, Micro-F1 score of 0.744, Macro Recall of 0.722, and Micro-Recall of 0.728, by employing binary relevance with logistic regression and BERT embedding. The use of binary relevance with SVM and Glove embedding yielded the highest Macro-Precision and Micro-Precision, reaching 0.944 and 0.933, respectively. Concerning the Zero-One Loss metric, the better outcome attained was 0.572 by using the Classifier Chain with Glove embeddings.