Large interest network for click-through rate prediction.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Recommender Systems

Background:

• Traditional recommender systems struggle with limited world knowledge and inefficient fine-grained interest modeling for click-through rate (CTR) prediction.
• Existing methods often fail to balance prediction accuracy with inference speed.

Purpose of the Study:

• To propose a unified framework, Large Interest Network (LIN), that addresses the limitations of current CTR prediction models.
• To leverage large language models (LLMs) for enhanced user and item profiling and improve interest modeling efficiency.

Main Methods:

• LIN integrates LLMs for generating semantically rich user and item profiles.
• Semantic Hard Negatives Contrastive Learning is employed to align different representation spaces.
• Cluster-based Interest Representation utilizes LLM profile cluster centroids for efficient and accurate modeling.

Main Results:

• LIN demonstrates consistent outperformance over state-of-the-art methods on multiple public datasets.
• Achieved AUC improvements of 0.46%-2.75% and LogLoss reductions of 7.52%-15.18%.
• LIN offers significantly faster inference speeds, 36×-244× quicker than interest-modeling approaches.

Conclusions:

• LIN effectively addresses key challenges in CTR prediction by incorporating external knowledge and optimizing interest modeling.
• The proposed framework achieves a superior balance between prediction accuracy and inference efficiency.
• LIN represents a significant advancement in developing practical and high-performing recommender systems.