Skip to content

Latest commit

 

History

History
20 lines (15 loc) · 2.23 KB

2402.10555.md

File metadata and controls

20 lines (15 loc) · 2.23 KB
Raw

Background

  • Background
    The paper discusses the importance of leveraging users' long-term engagement histories for personalized content recommendation but identifies existing challenges with processing extensive historical text and insufficient user-item interactions.

  • Existing Work Current research still struggles to encode very long user historical text and encounters problems with inadequate user-item interaction.

Core Contributions

  • The Content-Based Recommendation Framework SPAR

    • Challenge 1: Extracting holistic user interests from long engagement history The proposed SPAR framework leverages poly-attention layers and attention sparsity mechanisms in pre-trained language models (PLMs) to encode users' histories in a session-based manner, effectively tackling the challenge of extracting comprehensive user interests.

    • Challenge 2: Maintaining standalone representations for usability in practical deployment SPAR efficiently fuses user and item side features for engagement prediction while maintaining standalone representations, which is key for practical model deployment.

Implementation and Deployment

The results show that the SPAR framework outperforms our current state-of-the-art on two benchmark datasets, demonstrating its effectiveness. Ablation studies on the MIND-small dataset highlighted the significance of the UHS layer in summarizing user history, and the role of sparse attention in utilizing the UHS layer. The experiments indicated that removing the UIE layer, not using a PLM, keeping the PLM frozen, or removing session grouping, LLM-generated user-interest summaries, and the CCS layer resulted in slight performance decreases. However, removing all three components together led to a significant performance drop, underscoring their collective importance in the framework. The SPAR also showcased superior performance on various ranking metrics such as AUC when compared to other models that encode long sequences of user history content.

Summary

The SPAR framework effectively uses long-term user engagement histories to enhance the accuracy of personalized content recommendations and surpasses the existing state-of-the-art across multiple performance metrics.