Abstract
Sentiment Analysis (SA) of online product reviews plays a pivotal role in e-commerce by enabling
data-driven decision-making, product improvement, and personalized recommendation systems.
However, accurate and reliable sentiment classification remains challenging due to the massive
scale of user-generated reviews, multi-class label imbalance, and the limited interpretability of
machine learning models. To address these limitations, this study proposes a novel explainable
ensemble learning framework for multi-class SA (positive, neutral, and negative) across three
Amazon product domains (appliances, groceries, and clothing). The proposed study framework
integrates diverse supervised classifiers within a stacking ensemble, leveraging SHapley Additive
exPlanations (SHAP) explanations, innovatively employed not only to elucidate feature
contributions but also to rank the influence of individual base models on the ensemble predictions.
This approach represents a pioneering application in domain-specific SA, providing global and
local insights into model dynamics and base model selection. Unlike prior studies that relied on
Local Interpretable Model-agnostic Explanations (LIME), this framework enhances transparency
on the global and local levels while guiding ensemble optimization.
Evaluation using class-imbalance-sensitive metrics, including weighted and macro F1-scores,
Matthews Correlation Coefficient (MCC), Cohen's Kappa score, and Geometric Mean,
demonstrates that the stacking ensembles consistently outperform individual base models, with
Bernoulli Naïve Bayes excelling in recognizing the minority classes, as reflected by the Geometric
Mean. Across all three domains, the ensembles maintain strong and consistent performance,
highlighting their domain-agnostic capability to adapt to diverse product categories. SHAP
analyses further reveal domain-specific drivers and the roles of base classifiers, enhancing
transparency. The results underscore the framework's effectiveness in delivering both predictive
robustness and interpretability, enabling trustworthy automated sentiment analytics and informed
business decisions. This work establishes a foundation for future extensions to low-resource
categories, multimodal data, fine-grained sentiment scales, and enhanced class-imbalance
handling strategies, advancing the development of interpretable and scalable SA in e-commerce.