Online Item Cold-Start Recommendation with Popularity-Aware Meta-Learning

• DOI: 10.48550/arxiv.2411.11225
• Publication Date: 2024-11-17
• Authors: Yunze Luo; Yuezihan Jiang; Yinjie Jiang; Gaode Chen; Jingchi Wang; Kaigui Bian; Peiyi Li; Qi Zhang
• Keywords: Popularity; Cold start (automotive)

With the rise of e-commerce and short videos, online recommender systems that can capture users' interests update new items in real-time play an increasingly important role. In both offline recommendation, cold-start problem due to interaction sparsity has been affecting recommendation effect items, which is also known as long-tail item distribution. Many scheme based on fine-tuning or knowledge transferring shows excellent performance recommendation. Yet, these schemes are infeasible for streaming data pipelines different training method, computational overhead time constraints. Inspired by above questions, we propose a model-agnostic algorithm called Popularity-Aware Meta-learning (PAM), address under settings. PAM divides incoming into meta-learning tasks predefined popularity thresholds. The model distinguish reweight behavior-related features content-related each task their roles levels, thus adapting recommendations samples. These task-fixing design significantly reduces additional computation storage costs compared methods. Furthermore, introduced augmentation self-supervised loss specifically designed low-popularity tasks, leveraging insights from high-popularity This approach effectively mitigates issue inadequate supervision scarcity Experimental results across multiple public datasets demonstrate superiority our over other baseline methods addressing challenges scenarios.