Tourist Attractions Recommendation by using Enhanced Location Knowledge Graph

Location prediction has advanced through recommendation systems based on knowledge graphs, which excel in semantic and relational modeling. However, traditional knowledge graphs are primarily built around urban points of interest (POIs). Given that tourist activities outside the city exhibit unique patterns compared to daily urban behaviors, applying location-based knowledge graphs to tourist attraction prediction remains an ongoing research challenge. Furthermore, most studies focus on spatiotemporal data, often sidelining valuable textual information due to issues such as data scarcity and challenges in extracting meaningful semantics. To address these limitations, in our previous work, we introduced KGTR (Knowledge Graph with Tourist Reviews), a novel framework for predicting tourist attractions. Our framework goes beyond the traditional focus on temporal and spatial data by incorporating descriptive information on POIs visited by tourists, enhancing the knowledge graph’s expressiveness and semantic richness. Leveraging embeddings obtained through matrix factorization, we augmented the embeddings of locations and users with textual data to improve prediction accuracy. In this work, we compared three review embedding strategies: (a) before matrix factorization, (b) after matrix factorization, and (c) a combined before-and-after approach. To evaluate the performance, we collected a real-world tourism dataset containing POI user reviews. Results on both dense and sparse data indicate that the after factorization embedding consistently delivers the best performance. KGTR outperforms state-of-the-art methods, achieving an approximately 62.0% improvement in prediction accuracy. We further explored the impact of user clustering and found that finer cluster granularity enhances recommendation accuracy by better capturing user group preferences. Notably, the proposed after factorization embedding strategy remains the most effective. Additionally, we also compare KGTR with LightGCN, a widely used model in recommendation systems. Our experiments indicate that KGTR achieves the best performance, and by using text-enhanced information after matrix factorization, KGTR improves performance by 79.9% compared to LightGCN.