Deep fusion of multimodal features for social media retweet time prediction

摘要

The popularity of various social media platforms (e.g., Twitter, Facebook, Instagram, and Weibo) has led to the generation of millions of micro-blogs each day. Retweet (message forwarding function) is considered to be one of the most effective behavior for information propagation on social networks. The task of retweet behavior prediction has received much attention in recent years, such as modelling the followers' preference to predict if a tweet from others would be retweeted or not. But one important aspect in retweet behavior prediction is still being overlooked: the followers' retweet time prediction, which is helpful to understand the popularity of a tweet, the relationships between users, and the influence of users on their followers. However, due to the complex entanglement of multimodal features in social media such as text, social relationships, users' active time and many others, it is nontrivial to effectively predict the retweet time of followers. In this work, in order to predict the followers' retweet time on Twitter, we present an end-to-end deep learning model, namely DFMF (Deep Fusion of Multimodal Features), to implicitly learn the latent features and interactions of tweets, social relationships, and the posting time. Specifically, we adopt a word embedding layer to learn the high-level semantics of tweets and a node embedding layer to learn the hidden representations of the complex social relationships. Then, together with the one-hot representation of a tweet's posting time, the multimodal information is concatenated and fed into fully-connected forward neural networks for implicit cross-modality feature fusion, which is used to predict the retweet time. Finally, we evaluate the proposed method with a real-world Twitter dataset, the experimental results demonstrate that our proposed DFMF is more accurate in predicting the retweet time and can achieve as much as 11.25% performance improvement on the recall accuracy compared to Logistic Regression (LR) and Support Vector Machine (SVM).