AI Balances Fairness and Engagement in Video Recommendations

A new approach to AI-powered recommendation systems could help platforms like TikTok and KuaiShou deliver more engaging content while ensuring a fairer distribution of exposure across videos. Researchers have developed a that treats the lifecycle of a video—how its popularity changes over time—as a key factor in deciding what to recommend. This strategy aims to address a common problem where a few popular videos dominate user feeds, leaving newer or less-known content buried, which can stifle creativity and reduce user satisfaction over time.

The researchers found that short videos on these platforms follow a distinct three-phase lifecycle pattern, which differs from traditional models used for products or longer-form content. By analyzing real-world data from KuaiRec and KuaiRand datasets, they observed that videos typically experience rapid growth in engagement shortly after upload, followed by a brief period of stability, and then a sharp decline. This pattern is compressed, with over 90% of videos reaching peak activity within the first seven days. The study showed that user engagement, measured by play progress, varies significantly across these phases, with decline phases seeing up to a 43.3% drop compared to growth phases, highlighting the importance of timing recommendations to match a video's current stage.

To leverage this lifecycle insight, the team created a framework called LHRL, which stands for Lifecycle-aware Hierarchical Reinforcement Learning. This system uses a module called PhaseFormer to predict a video's current lifecycle stage in real time by analyzing its historical engagement data. PhaseFormer employs a technique called STL decomposition to break down play progress trends and seasonal patterns, feeding them into an attention-based model for accurate phase detection. The framework then operates with a two-level AI agent: a high-level agent sets fairness constraints based on the video's lifecycle stage, while a low-level agent optimizes immediate user engagement by recommending videos that balance these constraints with user interests.

Experiments on the KuaiRec and KuaiRand datasets demonstrated that LHRL outperforms 11 existing s in both user satisfaction and fairness. For instance, on KuaiRec, it achieved a 13.1% improvement in cumulative user reward and a 10% better fairness score compared to the best baseline. , detailed in Table 2 of the paper, show that LHRL maintains high interaction lengths and rewards while minimizing exposure disparities between popular and long-tail videos. Ablation studies confirmed that both the lifecycle-awareness and hierarchical structure are crucial, as removing either component led to significant performance drops, with flat agents struggling to balance long-term goals.

Of this research extend beyond short-video platforms to any interactive recommendation system where content popularity evolves over time. By dynamically adjusting exposure based on lifecycle stages, platforms can promote emerging content during critical growth phases, sustain engagement with mature videos, and reduce wasted exposure on declining items. This approach not only enhances user experience by keeping feeds fresh and relevant but also fosters a more equitable ecosystem where diverse creators have better opportunities to be discovered. 's generalizability was tested by integrating lifecycle-aware rewards into other AI models, yielding consistent performance gains, as shown in Table 4.

However, the study acknowledges limitations, such as the reliance on historical data for phase prediction, which may be sparse for new videos, though PhaseFormer uses extrapolation techniques to handle this. The framework also introduces moderate training overhead, and its performance depends on appropriately tuned hyperparameters, though the paper notes stability when these are emphasized. Future work could explore adapting the lifecycle model to other content types or platforms with different temporal dynamics, as the current analysis is focused on short-video datasets from specific time periods.