This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
In this paper, we propose a novel generative framework for multi-action recommender systems, Generative Recommendation via Journey-aware Sparse Attention on Chain-of-thought TokEnization (GRACE). To address the shortcomings of existing generative models, such as insufficient token inference information, high computational cost, and limited multi-scale modeling, GRACE proposes a hybrid Chain-of-Thought (CoT) tokenization method that leverages attribute information of product knowledge graph and Journey-Aware Sparse Attention (JSA) mechanism. CoT tokenization enables interpretable and action-relevant generation, and JSA selectively focuses on segments of compressed sequences to reduce computational cost. Experimental results show that GRACE significantly improves performance over existing state-of-the-art models while also reducing computational cost.
Takeaways, Limitations
•
Takeaways:
◦
We demonstrate that a hybrid CoT tokenization technique leveraging product knowledge graphs can be used to build interpretable and actionable generative models.
◦
Effectively reduce the computational cost of generative models through the JSA mechanism.
◦
Achieve performance gains that surpass existing best-performing models in multi-action recommender systems.
•
Limitations:
◦
There is a possibility that the performance improvement of the proposed model may be biased towards a specific dataset.
◦
Model performance can be significantly affected by the quality and completeness of the product knowledge graph.
◦
Further research is needed on the optimal parameter settings of the JSA mechanism.
