Not All Videos Become Outdated: Short-Video Recommendation by Learning to Deconfound Release Interval Bias (2408.17332v1)

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 However, not all videos become outdated; certain classic videos can still attract user's attention.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.

• The paper demonstrates that release interval acts as a confounder distorting recommendations and introduces the LDRI framework to correct this bias.
• It integrates a video recency sensitivity perceptron with traditional user-video matching models and applies backdoor adjustment for accurate causal inference.
• Experimental results on KuaiRand datasets show LDRI outperforms state-of-the-art models across TopK metrics such as RECALL@K, MAP@K, and NDCG@K, even in cold start scenarios.

Short-Video Recommendation by Learning to Deconfound Release Interval Bias

The paper "Not All Videos Become Outdated: Short-Video Recommendation by Learning to Deconfound Release Interval Bias," presented at the 18th ACM Conference on Recommender Systems (RecSys '24), addresses the pervasive issue of release interval bias in short-video recommender systems. Current systems often exhibit a bias towards recommending recently released videos, which can undermine the user experience by neglecting older yet still relevant videos. By identifying the release interval as a confounding factor, the authors propose a model-agnostic framework named Learning to Deconfound Release Interval Bias (LDRI) to mitigate this bias.

Fundamental Concepts and Methodology

The paper introduces the problem of release interval bias, substantiating it with empirical evidence from the KuaiRand-1K dataset. It demonstrates that the exposure and positive feedback rates for videos decrease with increasing release intervals. This bias can distort recommendation models, leading them to prioritize newer content over potentially more relevant older content.

The authors formalize the recommendation process using causal graphs, identifying the release interval as a confounder. This confounding effect creates non-causal relationships between video features and user interests, which the authors aim to eliminate through causal inference techniques, specifically backdoor adjustment.

Proposed Solution: LDRI Framework

The LDRI framework integrates a video recency sensitivity perceptron with traditional user-video matching models in a unified training setup. This perceptron captures the temporal sensitivity of each video, allowing for a nuanced understanding of how different videos age. The framework then applies backdoor adjustment during the inference stage to block the spurious paths introduced by the release interval confounder. This ensures that the resulting recommendation is based on the true causal effects of the video features on user interests, devoid of temporal bias.

Experimental Evaluation

Extensive experiments were conducted on two large-scale datasets: KuaiRand-Pure and KuaiRand-1K. The LDRI framework was compared against state-of-the-art recommender systems, including DeepFM, NFM, AFM, TCCM, and DCR-MoE. The evaluation demonstrated that LDRI consistently outperformed these models across various TopK recommendation metrics such as RECALL@K, MAP@K, NDCG@K, and HR@K.

One notable aspect of the evaluation was the framework's performance in cold start scenarios, particularly in recommending newly released videos. LDRI's ability to leverage both user feedback and intrinsic video features allows it to accurately model recency sensitivity even when user feedback is sparse. This adaptability is critical for maintaining recommendation quality in dynamic content environments.

Implications and Future Work

The findings indicate that addressing release interval bias can significantly improve the performance of short-video recommender systems. By learning the recency sensitivity of videos and applying causal inference techniques, the LDRI framework provides a more balanced and user-centric recommendation strategy. This approach not only enhances user satisfaction by surfacing relevant older videos but also promotes fairness for content creators by ensuring that quality content doesn't go unnoticed due to its age.

Moving forward, the authors suggest exploring the dynamic modeling of user interests over time. While LDRI adeptly captures video aging patterns, user preferences are also dynamic and can shift based on various factors, including recent interactions and emerging trends. Integrating temporal dynamics of user behavior with video recency models could further refine recommendation accuracy and relevance.

In summary, this paper offers a significant advancement in recommender system research by highlighting and addressing the overlooked issue of release interval bias. Through a robust causal framework, it sets a new precedent for designing recommendation systems that are not only temporally aware but also more aligned with user engagement patterns.