Prompt Tuning for Item Cold-start Recommendation

TL;DR

This work tackles item cold-start in recommender systems by introducing PROMO, a prompt-tuning framework that uses pinnacle positive feedback as prompt information and an item-specific personalized prompt network to address data-side gaps and model-bias issues. The authors show that pinnacle feedback aligns prompts with downstream tasks better than traditional content-based prompts and that per-item prompts reduce warm-item dominance while preserving critical signals. Two prompt-enhanced losses, L_pfpe and L_pape, alongside a simple fusion with a frozen backbone, enable efficient, parameter-light training. Extensive experiments on four public datasets and a production deployment demonstrate consistent improvements in cold-start recommendations and commercial metrics.

Abstract

The item cold-start problem is crucial for online recommender systems, as the success of the cold-start phase determines whether items can transition into popular ones. Prompt learning, a powerful technique used in natural language processing (NLP) to address zero- or few-shot problems, has been adapted for recommender systems to tackle similar challenges. However, existing methods typically rely on content-based properties or text descriptions for prompting, which we argue may be suboptimal for cold-start recommendations due to 1) semantic gaps with recommender tasks, 2) model bias caused by warm-up items contribute most of the positive feedback to the model, which is the core of the cold-start problem that hinders the recommender quality on cold-start items. We propose to leverage high-value positive feedback, termed pinnacle feedback as prompt information, to simultaneously resolve the above two problems. We experimentally prove that compared to the content description proposed in existing works, the positive feedback is more suitable to serve as prompt information by bridging the semantic gaps. Besides, we propose item-wise personalized prompt networks to encode pinnaclce feedback to relieve the model bias by the positive feedback dominance problem. Extensive experiments on four real-world datasets demonstrate the superiority of our model over state-of-the-art methods. Moreover, PROMO has been successfully deployed on a popular short-video sharing platform, a billion-user scale commercial short-video application, achieving remarkable performance gains across various commercial metrics within cold-start scenarios

Figures (5)

• Figure 1: The model accuracy on cold-start items with item-side taking different inputs for representations encoding. The positive feedback provides more task-relevant information as the model with positive feedback information gains higher accuracy on user preference.
• Figure 2: The histogram of prediction values on cold-start items and warm-start items. The X-axis represents the prediction value, and the Y-axis represents the sample amount of the corresponding range of scores.
• Figure 3: Overview of PROMO. PROMO utilizes item features and pinnacle feedback as the prompt information to generate prompt embedding and personalized prompt network, then PROMO optimizes the prompt embedding and eliminates the model bias by the pinnacle feedback prompt-enhance loss and the popularity-aware prompt-enhanced loss separately.
• Figure 4: The visualization of item representation involves the assignment of class labels to nodes through the application of K-means clustering on the original input data. The nodes are differentiated into various classes through the use of distinct colors. The representation of nodes is visualized utilizing t-SNE van2008visualizing.
• Figure 5: The memory retention rate corresponds with the accumulative negative feedback amount on MovieLens 1M dataset. Each bar represents the memory retention rate correponds to a certain negative feedback amount.