With the increasing commercial and social importance of Instagram in recent years, more researchers begin to take multimodal approaches to predict popular content on Instagram. However, existing popularity prediction approaches often reduce hashtags to simple features such as hashtag length or number of hashtags in a post, ignoring the structural and textual information that entangles between hashtags. In this paper, we propose a multimodal framework using post captions, image, hashtag network, and topic model to predict popular influencer posts in Taiwan. Specifically, the hashtag network is constructed as a homogenous graph using the co-occurrence relationship between hashtags, and we extract its structural information with GraphSAGE and semantic information with BERTopic. Finally, the prediction process is defined as a binary classification task (popular/unpopular) using neural networks. Our results show that the proposed framework incorporating hashtag network outperforms all baselines and unimodal models, while information captured from the hashtag network and topic model appears to be complementary.