In recent years, discussions surrounding sex education have gained considerable attention, as the lack of comprehensive sex education has been linked to various societal issues. While micro-video platforms offer new opportunities for disseminating sex education content, they have also contributed to the proliferation of sexually suggestive videos. Existing video classification methods face significant challenges in this context, such as the difficulty of abstract concepts, cross-domain variation, and training bias due to class imbalance. To address these challenges, we propose a method for classifying sexually suggestive videos. Our approach introduces a consensus-aware visual encoder to assist the model in focusing on the common features of videos within the same category at both the distribution and feature levels, while effectively filtering out irrelevant visual distractions. This improves the model’s ability to capture abstract and complex features. Additionally, we employ a label distribution-aware training strategy that allocates more learning capacity to tail classes, ensuring balanced learning across all categories. Experimental results on the SexTok dataset demonstrate that our method excels in classifying sexually suggestive videos, offering improved handling of abstract and imbalanced video content.