In order to achieve a deeper understanding of what hashtags are recommended by our proposed model, we offer a qualitative analysis of three case studies. We selected three representative types of micro-video (i.e., singing, sports and dance) in our dataset, and present their ground truth hashtags and the hashtags predicted by our proposed methods in Figure 6. From the frst example of singing scenario, we can see that the methods with user embedding module predicted more personalized hashtags (e.g., #bangerz tour and #rip hannah montana), which might bring by the knowledge coming from user’s previous posted video or hashtags. We have also notice a positive effect on the propagation mechanism on example (b) that two long-tail hashtags (e.g., #viral dance appears 77 times, and #kid dancer appears 68 times) are predicted by V2HT and V2HTw/oU with the hashtag propagation mechanism included.