Evaluation model of South Korean wellness tourism destination attractiveness based on deep learning and multimodal data

The increasing need for data-driven and interpretable destination assessment in smart tourism systems is addressed in this study by proposing an enhanced evaluation methodology for calculating the attractiveness of wellness tourist locations in South Korea. These methods fail to capture the nonlinear, multimodal, and relational nature of modern tourism perception data. To address the challenge, a proposed Deep Multimodal Wellness Tourism Attractiveness Evaluation Network (DMW-WTAEN) is a tightly integrated end-to-end deep learning architecture designed to model the progressive formation of wellness tourism destination attractiveness for heterogeneous data sources in South Korea. The proposed approach integrates features from spatio-temporal context, numerical operational indicators, destination images, and textual reviews in a synergistic manner. DMW-WTAEN employs the Semantic-Consistent Multimodal Alignment Transformer (SC-MAT) to align visual aesthetics and sentiment-rich textual embeddings, followed by a Graph-based Multimodal Dependency Reasoner (GMDR) that models latent interdependencies among destinations using attention-weighted relational learning. To enhance robustness and generalization under data sparsity, Uncertainty-Calibrated Contrastive Fusion (UCCF) is incorporated, enabling adaptive modality weighting based on confidence estimation rather than fixed fusion rules. The model uses self-supervised pretraining, contrastive cross-modal representation learning, and uncertainty-aware regression to calculate a continuous attractiveness score. According to quantitative results, DMW-WTAEN outperforms state-of-the-art baselines, including multimodal CNNLSTM, transformer-only fusion, and graph neural network models, achieving an RMSE reduction of 8.8%, a Spearman rank correlation of 11.75%, and a Kendall TA accuracy ranking of 14.45%. Ablation experiments also show that uncertainty-calibrated fusion outperforms deterministic fusion by 5.4%. A high-fidelity, interpretable, and scalable model for evaluating wellness tourism locations is provided. This model enhances the multimodal intelligence of tourism analytics and informs destination planners and policymakers.