We frst construct a graph to explore hashtag correlations with external knowledge, and then leverage existing structural knowledge to derive proper dependencies between frequent hashtags and long-tail hashtags. The propagation of such hashtag relation information is then used to modify the representation of initial hashtag representation. Afterwards, we utilize three parallel Long Short-Term Memory Networks (LSTMs) to model the sequential features for units in each modality and the outputs of the three LSTMs are projected into a common space. Finally, we employ an interactive embedding network to predict the interactions among hashtags, micro-videos, and users. The code is released here.