MultiModal Adaptive Fusion for Itinerary Generation Robust to Missing Modalities

Abstract

Tour route recommendation aims to generate feasible travel sequences based on user needs and time budgets. Most existing tour route generation models rely heavily on trajectory sequence information and underutilize multimodal data of attractions, leading to limited adaptability to real-world scenarios with missing modalities and a lack of explicit modeling of tourists’ diverse interests and visiting order preferences. To address these issues, this paper proposes a multimodal adaptive fusion method for tour route generation that is robust to missing modalities. It models user preferences through a gated fusion mechanism, multi-interest decoupling, and a recent-visit memory gating module, and generates feasible routes by incorporating pairwise order loss and beam search with duration penalty. Comparative experiments and ablation studies conducted on real-world tourism datasets verify the effectiveness of the proposed model.