Towards Reliable Customer Satisfaction Prediction: An AIML-Driven Multi-Modal Approach for E-Commerce Platforms

Abstract

E-commerce platforms generate massive amounts of user interaction data in the form of reviews, ratings, and purchase history. Accurate prediction of customer satisfaction from this multi-modal data is critical for improving user experience, enhancing personalization, and driving business growth. However, existing solutions suffer from several challenges, including the cold-start problem for new users and items, data sparsity in user–item interaction matrices, and the inability to combine multiple data modalities effectively. This proposes a novel AI/ML-driven multi-modal framework that integrates three complementary components: BERT-based textual embeddings for capturing the semantic and sentiment information in customer reviews, LightGCN-based graph embeddings for modeling collaborative user–item relationships and mitigating sparsity issues, and metadata feature encoders for incorporating user demographics, product attributes, and contextual signals. The outputs from these components are fused in a joint feature space and passed through a neural prediction layer to estimate customer satisfaction scores. The expected outcome is a robust, scalable, and explainable prediction system that achieves higher accuracy, handles cold-start scenarios effectively, and can be deployed as a real-time decision-support tool for e-commerce platforms through a Streamlit-based interface.