Not All Videos Become Outdated: Short-Video Recommendation by Learning to Deconfound Release Interval Bias

Lulu Dong, Guoxiu He, Aixin Sun

Introduction

Short-video platforms often suffer from a pronounced recency bias, where recommendation algorithms disproportionately favor newly released content. This bias stems from training on historical interaction logs where new videos received more exposure, leading models to conflate "newness" with "quality." However, content value is not uniform over time; while news clips expire quickly, "evergreen" content like tutorials or comedy can remain relevant for years.

The Challenge of Release Interval Bias

The authors identify the release interval—the time elapsed since a video's publication—as a statistical confounder. This confounder creates a "backdoor path" in the causal graph of recommendations, tricking models into over-promoting mediocre new clips while burying high-quality older ones.

Exposure distribution showing the heavy bias toward recently released videos.

As shown in the exposure distribution, the platform's existing mechanisms heavily prioritize videos with a release interval of 0-1 days, creating a feedback loop that reinforces this temporal bias. This makes it difficult for high-quality older content to compete for user attention.

Methodology: The LDRI Framework

To address this, the researchers proposed LDRI (Learning to Deconfound Release Interval Bias), a model-agnostic causal architecture. LDRI consists of three primary components:

  1. Matching Model: A standard backbone (e.g., DeepFM) that predicts user-video interest.
  2. Recency Sensitivity Perceptron: A component that learns how different content types (topics) decay over time based on inherent features.
  3. Backdoor Adjustment: Using do-calculus during inference to mathematically intervene on the release interval, effectively asking how a user would like a video regardless of its age.

Key Insights: Evergreen vs. Recency-Sensitive Content

A critical finding of the study is that different topics exhibit vastly different sensitivities to aging. By analyzing real-world data from Kuaishou, the authors observed that while some topics decay rapidly, others maintain or even increase their appeal over time.

Topic 2 demonstrates recency-insensitive behavior, where user interest (Favorite Rate) remains high or even grows despite decreasing exposure.

For instance, "Topic 2" (evergreen content) shows that user interest does not strictly follow the downward trend of exposure. In contrast, "Topic 20" (news/trending) shows a sharp decline in positive feedback as the release interval increases. LDRI leverages these differences to provide more nuanced recommendations by distinguishing between these decay curves.

Results and Impact

Extensive experiments on the KuaiRand datasets demonstrated that LDRI consistently outperforms state-of-the-art models across metrics like Recall, MAP, and NDCG. Notably:

  • Cold-Start Performance: LDRI improves recommendations for brand-new videos by predicting their recency sensitivity based on similar existing content.
  • Fairness for Creators: The model provides "fair opportunities" for older, high-quality videos to be rediscovered, extending the lifespan of valuable content.
  • Score Correction: Qualitative analysis showed that LDRI "smooths" the prediction scores for older videos, preventing the erratic fluctuations seen in standard models that rely too heavily on temporal shortcuts.

Conclusion

By treating the release interval as a confounder rather than a direct feature of quality, LDRI successfully deconfounds temporal bias in short-video recommendations. This causal approach ensures that "evergreen" content is not unfairly penalized for its age, leading to a more diverse and satisfying user experience.

Summary updated

Abstract

Short-video recommender systems often exhibit a biased preference to recently released videos. However, not all videos become outdated; certain classic videos can still attract user's attention. Such bias along temporal dimension can be further aggravated by the matching model between users and videos, because the model learns from preexisting interactions. From real data, we observe that different videos have varying sensitivities to recency in attracting users' attention. Our analysis, based on a causal graph modeling short-video recommendation, suggests that the release interval serves as a confounder, establishing a backdoor path between users and videos. To address this confounding effect, we propose a model-agnostic causal architecture called Learning to Deconfound the Release Interval Bias (LDRI). LDRI enables jointly learning of the matching model and the video recency sensitivity perceptron. In the inference stage, we apply a backdoor adjustment, effectively blocking the backdoor path by intervening on each video. Extensive experiments on two benchmarks demonstrate that LDRI consistently outperforms backbone models and exhibits superior performance against state-of-the-art models. Additional comprehensive analyses confirm the deconfounding capability of LDRI.
BibTeX
@inproceedings{dong2024not,
  author = {Lulu Dong and Guoxiu He and Aixin Sun},
  title = {{Not All Videos Become Outdated: Short-Video Recommendation by Learning to Deconfound Release Interval Bias}},
  booktitle = {RecSys},
  year = {2024},
  doi = {10.1145/3640457.3688113},
}
Figure