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

TL;DR

The paper tackles the item cold-start problem in online streaming recommender systems, where data are highly skewed toward popular items and real-time updates are needed. It introduces Popularity-Aware Meta-Learning (PAM), a model-agnostic framework that partitions data into fixed popularity-based tasks and learns task-specific parameters with a global meta-learning objective, while sharing knowledge across tasks. A cold-start task enhancer further augments learning through embedding simulations, data augmentation, and a self-supervised instructor that aligns cold-start embeddings with well-trained popular-item embeddings. Experiments on MovieLens, Yelp, and Book datasets, plus online A/B tests, show that PAM significantly improves cold-start item recommendations with competitive performance on popular items and reduced online computation and storage overhead. The approach offers a practical, scalable solution for real-time personalization in systems facing long-tail item distributions and streaming constraints.

Abstract

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

Figures (4)

• Figure 1: The overview of the proposed PAM method. $\Phi_t$ denotes the task-shared embedding parameters, $\Theta_t$ denotes the task-shared network parameters initialization that will be fine-tuned to specialized parameters $\Omega_t^n$ using $D_t^n$ for each task.
• Figure 2: The structure of our proposed cold-start task enhancer.
• Figure 3: Squared errors of top representations of cold-start items and popular items before and after masking for different types of embedding inputs.
• Figure 4: Performance of PAM w.r.t. different weights of cold-start tasks and different weights of losses.