The attraction recommendation systems not only filter out overwhelming irrelevant information for visitors but also identify potential customers for service providers. However, the current attraction recommendation methods such as content-based methods, collaborative filtering, or deep learning-based methods are either inaccurate due to data sparsity, or lack of interpretability, which results in the users suspicion on the recommendation results. To address the limitations of the current methods, we introduce a novel framework for preference propagation on knowledge graphs (KGs), which utilizes lots of parameters to capture the abundant semantics of existing KGs more comprehensively, and meanwhile explains the results through reasoning the link paths from users history to candidates on KGs. With a multi-view spatiotemporal analysis on real-world travel data, we investigate the geographical characteristics of human tour activities and build a tourism-oriented KG based on open web resources. Then, we propose a KG-aware attraction recommendation method named Geo-RippleNet and implement it with extensive experiments on large-scale datasets. It is argued that the framework for preference propagation on KGs not only absorb rich semantic information to achieve substantial performance gains in the attraction recommendation scenario but also enhance the interpretability of recommendation results with the support of abundant relational knowledge. Moreover, incorporating the spatiotemporal characteristics of human tour activities into the framework for preference propagation further makes the recommendation performance more aligned with the potential interests of visitors.