VEGAR: a tourist attraction recommender system based on signed feedback and a signed spatial-sentiment knowledge graph

Abstract

Tourism has grown rapidly. Travelers face abundant attractions and heterogeneous online content. This increases the cognitive effort of trip planning and motivates personalized recommendation. Existing recommender systems emphasize objective features but they exploit subjective review signals superficially, overlooking user personality cues and attraction-level signed sentiment. They also tend to treat users’ visits as uniformly positive feedback, even though negative preferences are common and informative in textual reviews. To address these issues, we propose VEGAR, a tourist attraction recommender system that jointly models signed feedback within a signed spatial-sentiment knowledge graph. To exploit subjective signals from reviews, VEGAR introduces a subjective feature extraction module and designs a signed spatial-sentiment knowledge graph construction framework. It extracts personality-related user features and attraction signed sentiment features from textual semantics. These subjective signals are organized together with functional features and spatial features in three subgraphs: a user subgraph, an attraction positive subgraph, and an attraction negative subgraph. To model signed feedback and align it with signed sentiment, VEGAR proposes a prediction module with five propagation procedures over three subgraphs to learn signed user and attraction preference profiles. VEGAR further introduces a high-order cross aggregation module and gated fusion to achieve informative user and attraction embeddings. To capture spatial constraints while preserving semantic relevance, VEGAR proposes spatial-sentiment attention that jointly models semantic relevance and distance-based influence during propagation. Experiments on three tourism datasets, namely the Suzhou, Beijing, and Alaska datasets, show that VEGAR achieves the best performance in both CTR prediction and Top-K recommendation. Additional ablation, robustness, cold-start, and interpretability analyses further validate the effectiveness and generalizability of the proposed framework.