Tourist behavior recognition and prediction based on a dual-stage attention fusion network

Tourist behavior recognition and prediction are core technologies for smart tourism systems, with significant importance for optimizing resource allocation in scenic areas and enhancing the tourist experience. However, existing methods exhibit clear shortcomings in handling multi-source heterogeneous data fusion and in modeling long-term sequential behavior. To this end, this article proposes a tourist behavior recognition and prediction model based on a dual-stage attention fusion network. This model incorporates three key contributions: an adaptive hierarchical attention-based fusion mechanism for multi-source heterogeneous data; a dual-stage attention network architecture that transitions from coarse-grained recognition to fine-grained prediction; and a behavior prediction framework. Specifically, the dual-stage architecture, through a cascaded design of spatio-temporally separated attention and semantically enhanced attention combined with a learnable gating mechanism, enables adaptive feature transfer between the two stages, capturing multi-scale spatio-temporal features of behavior sequences. Results show that our method improves accuracy by an average of 10.3% and reduces average displacement error by 27.6%, validating its effectiveness.