Predicting User Engagement in Live Video: A Method Based on Expectation Confirmation Theory

Abstract Existing research on live video recommendation primarily focuses on predicting the user's next click on a live video, neglecting the significance of user engagement. This paper proposes an algorithm based on Expectation Confirmation Theory (ECT) to predict user engagement by modeling user expectation and perceived experience. User expectation is determined by the historical experience of similar live videos, while the perceived experience depends on whether the target video aligns with the user’s evolving content preferences. To effectively model the evolution of user preferences, this paper employs a multi-head causal self-attention mechanism to capture user preferences and uses historical engagement sequences to control the mask matrix, capturing users' dynamic preferences. Finally, this paper integrates user expectation and perceived experience to predict engagement for each target video. To evaluate the performance of the proposed recommendation algorithm, experiments are conducted on a real-world live video dataset based on user's viewing behavior. The results demonstrate that the proposed algorithm outperforms baselines in both predicting user engagement and Top-N recommendation tasks. Moreover, this paper conducts several experiments to validate the robustness of the proposed recommendation algorithm and finally empirically tested the impact of experience and perceived experience on user engagement.