Modeling evolving user interests and engagement on short video sharing platforms: An attention-based deep generative approach

The rise of short video sharing platforms (SVSPs) has fundamentally transformed online content consumption. However, the unique characteristics of users’ short video consumption on SVSPs, including fine-grained temporal dependencies in rapid interest evolution, complex engagement state dynamics, and heterogeneous user–content attributes, pose significant challenges for understanding user behavior on SVSPs. To address these issues, we propose a Dynamic Interest and Engagement Model (DIEM), an attention-based deep generative model grounded in the Stimulus–Organism–Response (S–O–R) theoretical framework. Unlike conventional recommender models that represent users and content as static embeddings with heuristic interaction functions, DIEM models their interaction through a generative latent engagement process parameterized by a causal Transformer encoder and a bidirectional self-attention amortized inference network. By enforcing causal temporal structure in the encoder and leveraging bidirectional contextual information for posterior inference over engagement states, these architectures go beyond prevailing sequence models and effectively operationalize the “organism” component of S–O–R. We evaluated our model on the large-scale KuaiRand-1K dataset from Kuaishou platform. Experimental results demonstrate that DIEM significantly outperforms representative baselines across multiple evaluation metrics while providing interpretable insights into temporal interest evolution and personalized attention patterns. This research advances both theoretical understanding and practical applications for optimizing content delivery and enhancing user experience on SVSPs.