Dynamic Hashtag Assignment: Leveraging Graph Convolutional Networks with Class Incremental Learning

The use of hashtags plays a pivotal role in various applications. They have proven effective in social data mining, aiding information retrieval, sentiment analysis, event detection, and topic tracking. However, many users fail to include hash-tags, leaving a vast amount of content unnoticed. As a result, automating hashtag recommendations has become essential. This work introduces a novel class incremental learning approach for personalized hashtag recommendations using Graph Convolutional Networks (GCNs), leveraging image content and trending topics. In order to simulate the dynamic nature of social media trends and to validate the adaptability of our model to changing contexts, we create an extension of the popular HARRISON dataset by adding a temporal component. We investigate our solution's sensitivity to different approches and availability of training samples. The results presented show that our model can effectively adapt in different scenarios, whether old data is available or not at each training iteration, also through the use of different correlation matrices to mitigate computational and memory load. As far as we know, this work is the first incremental learning attempt at hashtag recommendation for real-world images in social networks. We expect this benchmark to accelerate the advancement of hashtag recommendation.