Long-tail Hashtag Recommendation For Micro-videos With Graph Convolutional Neural Network

Hashtag is one of the most iconic features on social media,which is used by users to tag,categorize and describe their uploaded content.In recent years,the micro-video industry has grown rapidly,with the traditional graphic form no longer satisfying consumers' content consumption needs,and the more intuitive micro-video medium becoming the most popular form of media.The stock of micro-videos on social media is already very large and still growing rapidly,which makes manual tagging become more and more impractical,and how to effectively organize and manage micro-videos has become a huge challenge for platforms.Micro-videos are usually associated with hashtags,which are often used to summarize the content of micro-videos and attract followers' attention.The use of hashtags can greatly improve the efficiency of microvideo retrieval,such as searching,browsing,and sorting,benefiting both users and administrators on the platforms.Despite the importance of hashtags,numerous microvideos lack hashtags or contain inaccurate or incomplete hashtags due to inconvenient input and other reasons.Therefore,personalized hashtag recommendation for useruploaded micro-videos has become an important research topic of common concern in academia and industry nowadays.The semantics of hashtags is consistent with the content uploaded by users,and it should be exactly the part that users are interested in.Personalized hashtag recommendation is to provide users with a list of suitable hashtags based on their uploaded content,which needs to take into account both the different preferences of users(different hashtags for posts with similar content)and the different habits of users in using hashtags(the same hashtags are used for different content).Previous work has focused on hashtag recommendation for microblogs or social images,and less research has been done on micro-videos.For the micro-video personalized hashtag recommendation task,micro-video feature modeling,user interest modeling,and hashtag semantic modeling are the key research problems to be addressed.There are three challenges for micro-video hashtag recommendation task:1)Temporal and Multimodal Characteristics of Micro-videos;micro-videos consist of visual,auditory and textual modalities,which need to consider temporal information,and joint multimodal representation.2)Diverse User Interests;user interests can be understood as user preferences for content and habits of using hashtags,which are complex and diverse,and need to obtain.3)Sparsity and Long-tail distribution of hashtags.At present,there is not enough research on the semantic aspects of hashtags,and how to solve the problem of long-tail distribution requires mining the semantics of hashtags.In this paper,we improve the performance of micro-video hashtag recommendation by proposing a novel graph convolutional neural network-based micro-video hashtag recommendation method by integrating multimodal joint representation for micro-video representation learning,user interest-based user representation learning,knowledge-guided hashtag representation learning,and collaborative interaction prediction model.By conducting experiments on real-world datasets we constructed,our proposed approach outperforms state-of-the-art benchmark methods.Meanwhile,we have released our dataset and code to facilitate research in this direction.