Wednesday 16 April 2025
The rise of online video platforms has led to a surge in user engagement, with comment sections becoming a crucial aspect of the viewing experience. However, predicting how users will interact with these comments remains a challenging task for recommender systems. A new study published in the ACM Transactions on Information Systems proposes a novel approach that leverages large language models (LLMs) to enhance comment staytime prediction.
The researchers behind this work collected a massive dataset from Kuaishou, a popular short-video platform, which consisted of over 10 million user interactions with comments. They analyzed various features, including video duration and watchtime, as well as user behavior such as likes and replies. By combining these factors with LLMs, the team developed a framework called Large Language Model-Enhanced Comment Understanding (LCU), which aims to improve staytime prediction accuracy.
The authors first fine-tuned the LLM on their dataset using a combination of domain-specific tasks, including comment understanding and ranking. This allowed the model to learn relevant contextual information about user behavior and video content. The LLM was then used as an embedding layer in a traditional recommendation model, which incorporated additional features such as video metadata and user profiles.
The results showed significant improvements in staytime prediction accuracy compared to baseline models that relied solely on traditional features. The LCU framework demonstrated an average relative improvement of 1.27% in predicting user engagement time, with notable gains for videos with shorter durations.
Further analysis revealed interesting patterns in the data. For example, as video duration increased up to around 600 seconds, staytime gradually increased, suggesting that longer videos encourage more engagement. However, beyond this point, the relationship between duration and staytime became less predictable. Similarly, watchtime had a significant impact on staytime, with users who watched longer portions of a video tending to spend more time in the comments section.
The authors acknowledge that their approach has its limitations, particularly in terms of scalability and generalizability to other platforms. However, they believe that LCU has the potential to become a valuable tool for recommender systems, enabling more accurate predictions and ultimately enhancing the user experience.
In the era of online video platforms, understanding how users interact with comments is crucial for optimizing the viewing experience. This study demonstrates the power of combining traditional features with large language models to improve staytime prediction accuracy.
Cite this article: “Enhancing Comment Staytime Prediction with Large Language Models”, The Science Archive, 2025.
Online Video Platforms, Recommender Systems, Comment Staytime Prediction, Large Language Models, Llms, Kuaishou, User Engagement, Video Duration, Watchtime, Comment Understanding, Ranking
