Hybrid Framework Addressing Imbalanced Data in Indonesian E-commerce Multimodal Sentiment Analysis

Severe class imbalance in Indonesian language e-commerce reviews hampers the detection of customer grievances, because about 88 percent of testimonials are positive whereas under one percent are negative. This study designs a simple yet effective framework to mitigate that bias. The data 2,611 authentic reviews covering electronics, fashion, and household goods and reflecting online shopping patterns were cleaned, structured, and represented using a combination of TF-IDF weighting and Word2Vec embeddings to condense meaning. Imbalance was addressed with the Synthetic Minority Oversampling Technique integrated with Edited Nearest Neighbour, equalising the proportions of positive, neutral, and negative classes. A Support Vector Machine was trained with five-fold cross-validation and benchmarked against Multinomial Naïve Bayes and a Decision Tree classifier. Experiments yielded 94.6 percent accuracy and a 94.8 percent F1-score, while precision for the negative class reached 100 percent, outperforming conventional approaches by up to twelve percent. The contribution of this research is the demonstration that minimal pre-processing, lightweight feature extraction, and targeted data balancing can markedly enhance grievance detection without resorting to complex model architectures. Practically, the framework enables online merchants to monitor service quality and respond to customer issues in real time, and it can be integrated into recommendation engines and business dashboards to accelerate data-driven decision making and strengthen transparency for stakeholders. Academically, the findings open avenues for integrating large language models, temporal analysis, and cross-domain adaptation to enrich the global e-commerce ecosystem.