MRLKG: Multimodal Representation Learning with Keyword-Based Graph for Hashtag Recommendation

The increasing volume of social media content has made hashtag recommendation a critical task for enhancing content discoverability and user engagement. This paper presents MRLKG (Multimodal Representation Learning with Keywordbased Graph) for hashtag recommendation. Our approach integrates cross-modality joint representation with a keyword-guided graph. The joint embedding provides a contextually rich representation of the content. In parallel, a keyword-guided Graph Convolutional Network (GCN) generates a refined representation of each microblog post content by emphasizing key concepts and their relational context with other microblog posts, essential for understanding the broader context and trends that hashtags often represent. By integrating the contextually rich joint embedding provided by transformers with the KeyBERT-GCN's focus on key concepts and their relational structure, the model can recommend hashtags that are both contextually relevant and aligned with the key trends represented by other microblog posts. Experimental results demonstrate the effectiveness of our method, showing improved performance over existing approaches. This method offers a robust solution for hashtag recommendation by leveraging both contextual understanding and graph-based reasoning.