A Multi-Branch Framework for Multidimensional Online Review Analytics

Information from the abstract

Online product reviews combine unstructured narratives, star ratings, and helpfulness votes, but these signals answer different managerial questions and require different validation strategies. This study presents a reproducible multi-branch analytical framework for the Amazon Fine Food Reviews dataset (568,454 records; October 1999–October 2012). The framework includes exploratory text mining, LDA topic modeling, rating-derived polarity classification, structured-feature score prediction, and rule-based rating-helpfulness segmentation. To address sampling bias and leakage, the revised experiments use reproducible random sampling, product-grouped holdout sets, and feature estimation only on training data. LDA models with K = 3–8 were compared on a random sample of 20,000 reviews across three seeds; K = 3 showed the best held-out perplexity and the strongest topic stability among the tested solutions. For rating-derived polarity, class-weighted TF-IDF Logistic Regression achieved 94.55% accuracy, 94.04% balanced accuracy, 90.47% macro-F1, 93.29% negative recall, and 93.27% negative-class PR-AUC on an independent product-grouped test set (n = 104,282). A regression tree using structural review features outperformed a mean predictor but remained modest (MAE = 0.884, RMSE = 1.194, R2 = 0.175), limiting its use to exploratory analysis. Helpfulness analyses treated the 270,052 reviews with no votes as Unrated rather than NotHelpful. Sensitivity analysis across thresholds of 0.50–0.80 and minimum-vote requirements of 1, 2, 5, and 10 showed that the estimated Helpful share depends materially on both choices. The revised evidence therefore supports descriptive screening and hypothesis generation, not causal claims or autonomous managerial decisions. The contribution is a transparent comparison of complementary analytical branches under leakage-aware evaluation, together with explicit construct, sampling, and external-validity limits. The findings show why class-sensitive metrics, topic stability checks, independent test sets, and separate treatment of unexposed reviews are necessary before review analytics can inform practice.

Why this record is monitored

This record has an Impact Signal of 73/100 based on recency, source, collaboration, and bibliographic signals. It prioritizes monitoring and is not a judgment of research quality.

Related topics: Sentiment Analysis and Opinion Mining · Computational and Text Analysis Methods · Topic Modeling