Multimodal Sentiment-Based Popularity Evaluation of Tourist Attractions Using Text, Image, and Geospatial Data Fusion

Abstract

This study proposes a novel tourist attraction popularity evaluation model that integrates multimodal sentiment analysis. The model combines a Transformer-based BERT network for textual sentiment classification, an improved ResNet convolutional network with a support vector machine (SVM) for image-based sentiment analysis, and Gaussian kernel-based spatial modeling for geographic heat estimation. Data collected from Zhangjiajie National Forest Park includes 300,000 user reviews, 50,000 images, and tourist origin data from 34 domestic provinces and 50 international countries. The proposed model achieved a text sentiment classification accuracy of up to 88%, an image sentiment F1-score of 0.82 in peak seasons, and a Pearson correlation coefficient of 0.92 between predicted heat values and actual tourist traffic. These results demonstrate strong predictive accuracy, cross-modal integration effectiveness, and robustness to noisy data, offering practical insights for attraction managers in real-time decision-making.