Engineering machine learning via integrating redundancy analysis in summarization with Non-negative Matrix Factorization and Latent Dirichlet Allocation

Abstract

The surge in e-commerce has generated vast volumes of unstructured customer reviews, making manual analysis inefficient and error-prone. Automated topic modeling methods often struggle to extract coherent product aspects while producing concise, non-redundant summaries. This study introduces the Enhanced Topic Modeling (ETM) framework, designed to integrate aspect extraction with redundancy-aware summarization for improved product review analysis. ETM leverages Latent Dirichlet Allocation (LDA) and Non-Negative Matrix Factorization (NMF) for aspect extraction, combined with both abstractive and extractive summarization techniques. A novel redundancy metric is proposed to measure and minimize repetition in summaries, balancing content richness with brevity. Performance evaluation shows that NMF outperforms LDA in aspect coherence (0.621 vs. 0.427). Abstractive summaries achieve higher conciseness but may lose detail, while extractive summaries retain more content but suffer higher redundancy. The ETM framework’s integration of aspect extraction, summarization, and redundancy evaluation enables more actionable insights from large-scale review datasets. By improving the clarity and usability of customer feedback analysis, ETM supports data-driven product development tailored to evolving consumer needs, addressing limitations of existing topic modeling approaches in the e-commerce domain. This approach advances both methodological rigor and practical application in customer sentiment intelligence.