Joint deep learning of sentiment and user profiles from electronic commerce reviews

E-commerce reviews provide a rich source of user generated content, capturing both fine grained emotional signals and latent user preference structures. Despite substantial progress in deep learning for sentiment analysis and user profiling individually, existing approaches treat these tasks in isolation by performing sentiment analysis as a preprocessing step and then feeding the outputs into downstream profiling or recommendation modules. This separation overlooks the intrinsic coupling between how users express emotions and who they are as consumers, limiting the expressive power of learned representations. A joint deep learning framework is proposed that simultaneously performs aspect based sentiment analysis and user preference modeling from e-commerce reviews. The framework employs a multi task architecture with shared Transformer based encoders, enabling bidirectional information flow between sentiment detection and user representation learning. Furthermore, we incorporate a fine grained emotion modeling component that extracts not only polarity but also specific emotion categories such as satisfaction, disappointment, and trust, grounded in psychological theory. Extensive experiments on two large scale e-commerce review datasets demonstrate that our joint framework achieves superior performance in both sentiment classification and user preference prediction tasks compared to task separated baselines, with particularly notable gains in cold start and data sparse scenarios. Ablation studies confirm that cross task synergy is the primary driver of performance improvement. The proposed framework establishes a foundation for truly integrated sentiment aware user profiling in next generation e-commerce personalization systems.