Warmer for Less: A Cost-Efficient Strategy for Cold-Start Recommendations at Pinterest

TL;DR

This work tackles the cold-start problem in large-scale recommender systems at Pinterest by introducing a cost-efficient, plug-and-play framework. It combines a residual path for non-historical features, a lightweight score debiasing loss based on MMD, and embedding-space manifold mixup to boost generalization to CS items without increasing serving costs. Offline and online experiments show consistent gains in fresh-content engagement, particularly for CS items, with limited parameter overhead and scalable deployment to hundreds of millions of users. The results demonstrate a practical path toward more balanced, generalizable recommendations in industrial-scale systems.

Abstract

Pinterest is a leading visual discovery platform where recommender systems (RecSys) are key to delivering relevant, engaging, and fresh content to our users. In this paper, we study the problem of improving RecSys model predictions for cold-start (CS) items, which appear infrequently in the training data. Although this problem is well-studied in academia, few studies have addressed its root causes effectively at the scale of a platform like Pinterest. By investigating live traffic data, we identified several challenges of the CS problem and developed a corresponding solution for each: First, industrial-scale RecSys models must operate under tight computational constraints. Since CS items are a minority, any related improvements must be highly cost-efficient. To address this, our solutions were designed to be lightweight, collectively increasing the total parameters by only 5%. Second, CS items are represented only by non-historical (e.g., content or attribute) features, which models often treat as less important. To elevate their significance, we introduce a residual connection for the non-historical features. Third, CS items tend to receive lower prediction scores compared to non-CS items, reducing their likelihood of being surfaced. We mitigate this by incorporating a score regularization term into the model. Fourth, the labels associated with CS items are sparse, making it difficult for the model to learn from them. We apply the manifold mixup technique to address this data sparsity. Implemented together, our methods increased fresh content engagement at Pinterest by 10% without negatively impacting overall engagement and cost, and have been deployed to serve over 570 million users on Pinterest.

Figures (7)

• Figure 1: The overview of key challenges in CS recommendation and how the proposed framework addresses them.
• Figure 2: (a) $\Delta$PR-AUC ratio from ablating historical and non-historical features, indicating stronger dependence on historical signals, with the ratio staying well below 100%. (b) Ratio of average $\ell_2$-norm of prediction gradients with respect to historical and non-historical features during training, illustrating that model updates are dominated by historical signals, as the ratio remains below one. (c) Visualization of predicted scores for randomly selected warm and cold samples (both positive and negative samples), illustrating that the model assigns substantially higher scores to warm items. This prediction bias leads to recommendations dominated by historical-rich items and systematic under-recommendation of cold items.
• Figure 3: (a) A residual connection to enhance the influence of non-historical features on model predictions. (b) A debiasing loss to reduce model bias toward warm instances, without requiring sampling techniques or computational overhead to the training. (c) Illustration of the mixup procedure: for each sample in a batch, a distinct random sample is chosen, and a new instance is generated via linear interpolation.
• Figure 4: Related Pins Surface on Pinterest. (a) The Pin, i.e., query Pin, user just grid-clicked on from an upper-stream surface. (b) Grid view of recommended Pins. User name and profile are masked out for privacy protection.
• Figure 5: Multi-task Ranking Model Architecture in Pinterest Related Pins Recommender System