Evaluating Difference-Based Dynamic Features for E-Commerce Content-Related Sentiment Classification Using a Fine-Tuned LightGBM Model

The increasing reliance on e-commerce services has generated extensive product-related information and customer opinions, enabling analysis of product-page evolution and consumer sentiment. However, repeated product-page observations do not directly represent temporal changes without extracting differences between snapshots. This study proposes a difference-based dynamic feature extraction methodology using time-normalized numerical differences and TF-IDF cosine distance. The extracted features were integrated with original product-related features and evaluated using tuned LightGBM to predict the dominant Fine-Grained Consumer Sentiment related to the Product (FG-CSrP) class at later observations of previously observed products. Performance was assessed through expanding temporal validation windows and an independent temporal holdout. LightGBM gain-based and group-level permutation importance were further used to examine feature contributions. On the final holdout, Baseline and Original + Dynamic settings achieved Macro-F1 scores of 0.8710 and 0.8616, respectively, showing no clear improvement from dynamic features. Additional benchmark and robustness analyses highlighted temporal stability, limited generalization to unseen products, and superior Macro-F1 performance of LightGBM compared with regularized logistic regression. LightGBM gain-based and group-level permutation importance analyses showed that original product-related features contributed most to the predictive performance. Overall, difference-based features explicitly represent changes in evolving e-commerce content, although these features did not demonstrate additional predictive gains.