Muhammad Syafiq, Wawan Saputra, Carlya Agmis Aimandiga, Cici Suhaeni, Bagus Sartono, Gerry Alfa Dito

Making Sense of Fashion Feedback : Comparing Two Popular Text Analysis Tools

Introduction

Making sense of fashion feedback : comparing two popular text analysis tools. Compare Word2Vec & GloVe for fashion review analysis using ML models. This study analyzes sentiment & predicts recommendations from 23k reviews, finding Word2Vec-LSTM achieved 87.35% accuracy.

Abstract

The rapid expansion of the fashion industry, propelled by digital technology and e-commerce, has resulted in a significant volume of customer-generated reviews. These reviews serve as a valuable source for understanding customer satisfaction and behavior. This study aims to (1) analyze customer sentiment, (2) predict product recommendations, and (3) examine the relationship between sentiment classification and recommendation decisions using text embeddings from Word2Vec and GloVe. The research utilized over 23,000 fashion product reviews sourced from Kaggle. Text data were preprocessed and vectorized using Word2Vec and GloVe, followed by classification and prediction tasks using six machine learning models: Random Forest, SVM, Naïve Bayes, LSTM, Logistic Regression, and Gradient Boosting. The results revealed that Word2Vec consistently outperformed GloVe across all models and tasks, with the Word2Vec-LSTM combination achieving the highest accuracy of 87.35% and F1 score of 92.35% in imbalanced data scenarios. Correlation analysis also confirmed a strong and statistically significant relationship between sentiment and recommendation labels, with Spearman’s Rho of 0.8340 and Kendall’s Tau of 0.8120. These findings suggest that high-quality sentiment representation can effectively support product recommendation systems. This study contributes to the understanding of embedding effectiveness in fashion-related text analysis and opens avenues for hybrid and transformer-based representations in future research.

Review

This paper, titled 'Making Sense of Fashion Feedback: Comparing Two Popular Text Analysis Tools,' addresses a timely and relevant challenge: leveraging the vast amount of customer-generated reviews in the burgeoning digital fashion industry. The study adeptly sets out to achieve three clear objectives: analyzing customer sentiment, predicting product recommendations, and elucidating the relationship between these two aspects. Employing a robust methodology, the authors utilize a substantial dataset of over 23,000 fashion product reviews from Kaggle. The core approach involves preprocessing and vectorizing text data using two widely recognized word embedding techniques, Word2Vec and GloVe, subsequently testing their efficacy across a diverse suite of six machine learning models, including traditional classifiers (Random Forest, SVM, Naïve Bayes, Logistic Regression, Gradient Boosting) and a deep learning model (LSTM). This comprehensive comparative framework is a significant strength, promising valuable insights into the performance of these tools in a specific domain. The findings presented are compelling and offer clear contributions to the field. A central discovery is the consistent superior performance of Word2Vec over GloVe across all evaluated models and tasks. Notably, the Word2Vec-LSTM combination emerged as the most effective, achieving impressive metrics of 87.35% accuracy and an F1 score of 92.35% in the context of imbalanced data. Beyond prediction, the study establishes a strong and statistically significant correlation between sentiment classification and recommendation decisions, quantified by high Spearman’s Rho (0.8340) and Kendall’s Tau (0.8120) values. These results collectively underscore the critical role of high-quality sentiment representation in building effective product recommendation systems within the fashion domain. The paper thus provides a clear empirical demonstration of embedding effectiveness and their practical utility. While the study offers valuable insights, a few aspects warrant further consideration. The abstract mentions "imbalanced data scenarios," and while the F1 score suggests some robustness, a more explicit discussion on how this imbalance was addressed (e.g., sampling techniques, cost-sensitive learning) would enhance the methodological clarity and confidence in the reported metrics. Furthermore, while the comparison between Word2Vec and GloVe is thorough, the rationale for selecting these specific six machine learning models, particularly the mix of traditional and one deep learning model, could be elaborated for a more complete picture. The paper aptly points to avenues for future research, suggesting hybrid and transformer-based representations. Expanding on potential limitations of the current embedding models in capturing nuanced fashion feedback (e.g., sarcasm, domain-specific terminology beyond generic embeddings) could provide a stronger motivation for these future directions. Nonetheless, this work lays a solid foundation, offering a pragmatic evaluation of current text analysis tools for fashion feedback, which is highly beneficial for practitioners and researchers alike.

Full Text

Comments

You need to be logged in to post a comment.