September 2023
·
20 Reads
·
1 Citation
The prediction of the popularity of online content, particularly videos, has recently gained significant attention as successful popularity prediction can assist many practical applications such as recommendation systems and proactive caching, as well as aid in optimizing advertising strategies or balancing network throughput. Despite much work being done on predicting the popularity of online videos, there are still challenges to be overcome: (1) popularity is greatly influenced by various external factors, resulting in significant fluctuations that are difficult to capture and track; (2) online video content and metadata information are typically diverse, sparse, and noisy, making the prediction task complex and unstable; (3) some data have temporal relevance, and the impact on popularity varies at different times. In this paper, we propose an Adaptive Temporal Knowledge Graph Network (ATKN) video popularity prediction model to address the issues surrounding video popularity prediction. First, we employ the attention-based Long Short-Term Memory (ALSTM) network to capture the trend of popularity change. Then, we introduce an Attention-based Factorization Machine (AFM) with attention mechanism to model the feature cross of video content, thereby enhancing the distinction of importance after different feature crosses. Next, we use a Relational Graph Convolutional Network (RGCN) to extract the associated features between entities in the knowledge graph. Finally, we propose a dynamic feature fusion method that adaptively assigns the weights of temporal features and content features at different time intervals by constructing an exponential decay function, thereby obtaining an effective and stable feature fusion module. Experimental results demonstrate the superiority and interpretability of ATKN on the MovieLens-20M dataset and the Microsoft Satori-built movie knowledge graph.