Not All Videos Become Outdated: Short-Video Recommendation by Learning to Deconfound Release Interval Bias

TL;DR

This work proposes a model-agnostic causal architecture called Learning to Deconfound the Release Interval Bias (LDRI), which enables jointly learning of the matching model and the video recency sensitivity perceptron, and demonstrates that it consistently outperforms backbone models and exhibits superior performance against state-of-the-art models.

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 However, not all videos become outdated; certain classic videos can still attract user's attention.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

Figures (7)

• Figure 1: (a) and (b) plot video exposure rate and user positive feedbacks against video release interval in days, on all records of KuaiRand-1K. (c) and (d) show the trend of user positive feedback for videos on two different topics with similar exposure distribution. Watch-time rate is the ratio of the watch time to the video length; favorite rate represents the ratio of users who liked a video to those who watched it.
• Figure 2: Causal view of short-video recommendations. These causal graphs represent traditional recommendation, ideal recommendation incorporating recency sensitivity of video, recommendation confounded by release interval bias, and decoufounded recommendation through backdoor adjustment, respectively. Nodes in the graphs are: video features ($V$), user features ($U$), matching between video and user ($M$), user interests ($UI$), recency sensitivity of video ($T$), and release interval confounder ($A$).
• Figure 3: Overview of the proposed LDRI framework. The left and middle parts illustrate the joint training of matching model between user features and video features, as well as perceptron of recency sensitivity in training stage. The right part depicts the inference stage, where LDRI fuses user interests and recency sensitivity of video, and implements backdoor adjustment to mitigate the confounding effect introduced by release interval.
• Figure 4: NDCG@5 results at each release interval on KuaiRand-Pure with three backbones and the enhanced version with LDRI. Similar observations hold with NDCG@300 on KuaiRand-1K.
• Figure 5: Metric@300 on newly released videos from KuaiRand-1K. The right four bars represent the average results of three backbones, backbones enhanced with TaFR, DCR-MoR, LDRI, respectively. Compared to backbones, the improvements by LDRI on all four metrics are significant with $p<0.05$.