A dynamic next-stop or next-item recommendation system that is built entirely from raw card transaction data logs. These data logs contain rich transaction data between cardholders and merchants. A query network approach is constructed for geometrical expressivity and automatically learns the inherent class-hierarchy. To ensure scalability and interpretability of the approach, merchants or entities are grouped into interpretable categories and propose a quadtree-based spatial decomposition of the underlying geography. A two-step recommendation process initiates: (1) predict next-merchant quadtree-box and category combination (2) recommend merchants within the predicted combination. This novel neural architecture may handle the hierarchical classification task in the first part of the recommendation system and compare the methods to previous state-of-the-art approaches in related areas.