Social media platforms generate an overwhelming volume of opinions and reviews that influence markets, politics, and consumer decisions. While today much attention is directed toward large language models (LLMs), this research explores non-LLM, feature-driven approaches for sentiment analysis and recommendations.

We focus on polarity detection—identifying whether a text expresses a positive, negative, or neutral sentiment—using Twitter/X data and sentiment lexicons. Novel feature modeling techniques and feature fusion strategies are introduced to improve sentiment classification and clustering of opinions.

Beyond analysis, we integrate sentiment detection into Recommendation Systems. A new similarity measure, ARP (Attractiveness–Relevance–Popularity), enhances hotel and restaurant reviews with sentiment information to generate more meaningful suggestions.

Experiments show that carefully engineered features and lexicon-based methods remain powerful alternatives to black-box LLMs. Our results highlight that sentiment-aware recommendations and interpretable feature design provide actionable insights for businesses, marketers, and users navigating the ever-growing stream of online opinions.