Complementary Model Fusion for Enhanced Social Popularity Prediction

Predicting the popularity of social media posts is an important and challenging task. However, existing approaches have limitations in effectively modeling the multimodal nature of posts and fusing diverse predictive signals. This paper proposes a novel deep neural network framework that integrates comprehensive features across textual, visual, social, and temporal modalities to predict photo popularity on Twitter and Flickr. The model employs parallel convolutional neural network streams to separately learn representations from visual and textual features. A late fusion scheme is introduced to combine these representations in a joint embedding space. Additionally, self-attention mechanisms are incorporated to focus on the most informative signals for popularity prediction. Through extensive experiments on large-scale Twitter and Flickr datasets, the proposed approach is shown to significantly outperform existing methods relying on individual modalities. Ablation studies validate the contributions of the key model components including multimodal feature extraction, late fusion, and self-attention. The results highlight the importance of integrated modeling with joint multimodal learning for accurately predicting social media virality. The techniques presented provide an effective and general framework for popularity prediction tasks involving diverse heterogeneous data. This work represents a significant advance in harnessing multimedia analytics and neural multimodal fusion to achieve state-of-the-art performance in predicting social media popularity.