Benchmarking Model Transformer Modern untuk Analisis Sentimen dan Tren Konsumen dalam Industri Fashion

Abstrak

The dynamic trends and consumer preferences in the fashion industry demand accurate sentiment analysis to support decision-making. This study aims to compare the performance of four Transformer models BERT, DistilBERT, RoBERTa, and DeBERTa, in sentiment classification of fashion product reviews to identify which model best captures the semantic nuances specific to this domain. Using an experimental quantitative approach, we analyzed 278,098 TeePublic reviews that underwent text cleaning, class balancing, and tokenization, and were then split into training, validation, and test sets. The four models were fine-tuned using Optuna-optimized hyperparameters and evaluated with accuracy and F1-score. Our results show that all models achieved nearly identical performance, with F1-scores and accuracy ranging from 0.824 to 0.826. ANOVA testing confirmed no significant differences among models, indicating that increased architectural complexity does not yield substantial benefits for a task of this scale. These findings suggest that performance consistency is driven more by preprocessing quality, class-balancing strategies, and pipeline design. The study offers a comparative basis for model selection in fashion-domain sentiment analysis and highlights that optimizing the processing pipeline often brings more meaningful improvements than adopting larger architectures.

Identifier

• OpenAlex: https://openalex.org/W7124720451
• DOI: https://doi.org/10.29408/edumatic.v9i3.32657