Research On Hashtag Recommendation Algorithm For Multi-Modal Data In Social Networks

Hashtag is a topic tag with a special form.In social networks,it is often used to identify the topic of the content posted by users.In view of the wide application of Hashtag and the huge data scale of social network platforms,more and more social network platforms begin to retrieve and manage the data on the platform through Hashtag.Due to the unreasonable Hashtag annotation of a part of the data,an effective Hashtag recommendation method is needed to recommend reasonable Hashtag tags for users.Multi-modal data understanding and user personalized recommendation are two key issues to be considered in the research of Hashtag recommendation methods.One of them,the understanding of multimodal data refers to how to establish the relationship model between different modal data,so as to effectively integrate the feature vectors of different modal data,and realize the effective extraction of the features of multimodal data.User personalized recommendation refers to how to recommend the tags that conform to the user's markup habits according to the user's markup history information.In the aspect of multimodal data understanding,the existing methods usually adopt a two-stage processing approach.In the first stage,the alternating attention mechanism is used to extract the feature vectors of each modal data,and in the second stage,the sum of the feature vectors of different modal data is used to calculate the final feature vectors.As the alternating attention mechanism only focuses on the correlation information between the data,the fusion method of feature vectors based on simple summation does not consider the semantic differences among the feature vectors of different modal data.Therefore,the existing methods have the problems of missing discrepant information and semantic bias.For that reason,a new feature extraction method based on graph attention mechanism is proposed to understand multimodal data.The proposed method integrates the correlation information and the difference information of image and text features by establishing three types of relationship models: image to image,image to text,and text to text,and uses graph attention mechanism to reduce semantic differences in multimodal data.In the aspect of user personalized recommendation,the existing methods first randomly sample the user history data.Then,word embedding method and attention mechanism are used to extract features from the sampled Hashtag sequences.The similarity between the target data and the user's historical data is calculated to obtain the weight of the Hashtag sequence feature vectors corresponding to each user's historical data,so as to realize the Hashtag recommendation based on the user's tag behavior habits.Since random sampling method cannot filter noise information and Hashtag sequence feature extraction method is difficult to achieve deep semantic understanding,existing methods cannot accurately depict users' markup habits.Therefore,this paper studies the problem of user personalized recommendation,and proposes a personalized recommendation method based on memory unit.Different from the existing methods,the proposed method only samples the user history data with high cosine similarity to the target data,and uses Transformer encoder structure and attention mechanism to extract the features of Hashtag sequences from the samples.Finally,combined with the proposed multi-modal data understanding method and user personalized recommendation method,a Hashtag recommendation model that supports multi-modal data and considers user markup habits is constructed.The test results on Twitter and Instagram datasets containing image and text data show that the proposed feature extraction method based on graph attention mechanism can understand the semantics of multimodal data more accurately than the existing methods.Tests on an Instagram dataset containing historical tagged data show that the Hashtag recommendations of the proposed model outperform existing methods.