Micro videos popularity prediction (MVPP) has recently attracted widespread research interests given the increasing prevalence of video-based social platforms. However, previous studies have overlooked the unique patterns between popular and unpopular videos and the interactions between asynchronous features different data dimensions. To address this, we propose a novel hierarchical attention contrastive learning method named HACL, which extracts explainable representation features, learns their asynchronous interactions from both temporal and spatial levels, and separates the positive and negative embeddings identities. This reveals video popularity in a contrastive and interrelated view, and thus can be responsible for a better MVPP. Dual neural networks account for separate positive and negative patterns via contrastive learning. To obtain the temporal-wise interaction coefficients, we propose a Hadamard-product based attention approach to optimize the trainable attention-map matrices. Results from our experiments on a TikTok micro video dataset show that HACL outperforms benchmarks and provides insightful managerial implications.


Paper Number 1221; Track Design; Complete Paper



When commenting on articles, please be friendly, welcoming, respectful and abide by the AIS eLibrary Discussion Thread Code of Conduct posted here.