UniShare: A Unified Framework for Joint Video and Receiver Recommendation in Social Sharing

TL;DR

UniShare presents a unified framework for jointly predicting video share probability and receiver recommendations by integrating enhanced bilateral representations, multi-modal signals, and relationship-content alignment. The approach employs a joint training paradigm with hierarchical negative sampling and cross-task parameter sharing to mitigate data sparsity and boost both tasks, validated on the large-scale K-Share dataset. Offline evaluations show clear gains over strong baselines, and online A/B tests on Kuaishou demonstrate significant improvements in shares, unique sharers, and receiver engagement. The work also releases K-Share and demonstrates practical deployment considerations, highlighting the approach's real-world impact for social sharing ecosystems.

Abstract

Sharing behavior on short-video platforms constitutes a complex ternary interaction among the user (sharer), the video (content), and the receiver. Traditional industrial solutions often decouple this into two independent tasks: video recommendation (predicting share probability) and receiver recommendation (predicting whom to share with), leading to suboptimal performance due to isolated modeling and inadequate information utilization. To address this, we propose UniShare, a novel unified framework for joint sharing prediction on both video and receiver recommendation. UniShare models the share probability through an enhanced representation learning module that incorporates pre-trained GNN and multi-modal embeddings, alongside explicit bilateral interest and relationship matching. A key innovation is our joint training paradigm, which leverages signals from both tasks to mutually enhance each other, mitigating data sparsity and improving bilateral satisfaction. We also introduce K-Share, a large-scale real-world dataset constructed from Kuaishou platform logs to support research in this domain. Extensive offline experiments demonstrate that UniShare significantly outperforms strong baselines on both tasks. Furthermore, online A/B testing on the Kuaishou platform confirms its effectiveness, achieving significant improvements in key metrics including the number of shares (+1.95%) and receiver reply rate (+0.482%).

Figures (6)

• Figure 1: A screenshot illustrating the sharing flow on the Kuaishou platform. The process consists of three stages: (1) watching a video and generating sharing intent and clicking the share button to invoke the sharing panel, and (2) selecting a receiver to complete the sharing action, followed by (3) the receiver viewing the video and providing feedbacks.
• Figure 2: Overview of the UniShare Framework. UniShare jointly trains the video recommendation and receiver recommendation tasks, enhancing performance on both through underlying information sharing. Pre-computed GNN embeddings, multi-modal embeddings, and relationship alignment features are incorporated as enhanced inputs to the model.
• Figure 3: Comparison of PCOC change when modifying the input candidate set $\mathcal{V}_u$. $\mathcal{V}_u$-ctrl denotes the removal of the most frequently shared receiver from the candidate set.
• Figure 4: Example illustration of the effects of RCA and BIM
• Figure 5: For a single user request, a joint recommendation approach requires estimating a subsequent receivers ranking for every candidate video. We retrieve a subset $\mathcal{\hat{V}}_u$ from $\mathcal{V}_u$ for UniShare. The final output is produced by merging and re-ranking the results from this subset with those from the full candidate set.