Text analysis and location recommendation on location based social networks (LBSNS)

Purpose: The purpose of this thesis is to enhance recommendation systems by leveraging Heterogeneous Information Networks (HINs) integrated with Aspect-Based Sentiment Analysis (ABSA). It aims to address challenges in capturing nuanced user-business relationships and improving recommendation accuracy through review-enriched meta-path-driven embeddings. Two innovative frameworks, REHREC and W-REHREC (Weighted REHREC), are proposed to incorporate semantic and structural insights from reviews and interactions. The study also evaluates the adaptability of these models across diverse datasets, including review-rich environments (Yelp, Foursquare). This research contributes to bridging the gap between textual and structural data while addressing cold-start issues, sparse interactions, and the need for scalable and generalizable recommendation methods. Methods: The proposed REHREC and W-REHREC frameworks utilize meta-path-driven HIN embeddings to capture direct and indirect relationships among users, businesses, and reviews. W-REHREC incorporates edge weights derived from ABSA sentiment scores, prioritizing impactful interactions. Experiments were conducted on two datasets: Yelp and iv Foursquare, each representing varying data characteristics. Evaluation metrics, including Root Mean Square Error (RMSE) and Mean Absolute Error (MAE), assessed recommendation accuracy. Meta-paths such as UBU, BUB, URRU, BRRB, UBRRBU, UBCaBU, and BCaB were utilized to model complex relationships between users (U), businesses (B), reviews (R) and categories (Ca). The models were trained with random walks and skip-gram embeddings, with hyperparameters tuned for optimal performance across datasets. Results: The proposed frameworks demonstrated significant improvements in recommendation accuracy across all datasets. On the review-rich Yelp dataset, W-REHREC achieved an RMSE reduction of 15.65% compared to the baseline HERec model, with similar improvements observed for MAE. On the Foursquare dataset, W-REHREC improved RMSE by 14.11%, showcasing its generalization capability. Meta-paths capturing semantic and structural relationships were instrumental in these improvements, with W-REHREC consistently outperforming REHREC by emphasizing sentiment-driven connections. Conclusion: This thesis presents a robust contribution to recommendation systems through the development of REHREC and W-REHREC frameworks. By integrating ABSA with HIN embeddings, these models effectively leverage review data to capture nuanced user-business relationships. W-REHREC, with its sentiment-weighted edges, significantly enhances accuracy and scalability, demonstrating adaptability across diverse datasets. The results highlight the importance of combining semantic and structural insights for modern recommendation systems. This work provides a foundation for future research in multi-language sentiment analysis, dynamic temporal modeling, and real-time recommendation deployment, addressing evolving challenges in personalized, context-aware, and scalable recommendation technologies. Keywords: Recommendation Systems, Heterogeneous Information Networks, Location Based Social Networks, Aspect-Based Sentiment Analysis, Meta-Path Embeddings, Collaborative Filtering.