Search papers, labs, and topics across Lattice.
This paper introduces TSPORec, a novel approach for optimizing token selection in LLM-based sequential recommendation systems, addressing the inefficiencies of existing methods that rely on truncated item descriptions. By implementing a three-stage pipeline to identify and utilize informative tokens from full text, TSPORec significantly enhances both recommendation performance and computational efficiency. Experimental results show that TSPORec achieves up to 31.25% improvement in recommendation accuracy and up to 63.4% reduction in inference costs compared to six baseline methods.
TSPORec reveals that leveraging full-text information can boost recommendation accuracy by over 31% while slashing inference costs by more than 63%.
Large Language Models (LLMs) have emerged as powerful tools for improving recommendation systems. The effectiveness of LLMs arises from their ability to harness rich textual information and their capacity to model heterogeneous user preferences based on users'interaction history. However, due to the large-scale and deep architectures, LLM-based sequential recommendation approaches generally incur high inference costs, resulting in a low return on investment. To mitigate this cost, many existing approaches resort to using only the first few tokens of item descriptions, which inadvertently discards valuable information contained in the full text, thereby leading to suboptimal recommendation performance. To address this limitation, we propose a novel Token Selection approach for Preference Optimization in LLM-based sequential Recommendation, i.e., TSPORec, which accurately pinpoints informative tokens throughout the entire textual content to improve recommendation performance. Specifically, we design a three-stage pipeline to select informative tokens and introduce a novel proxy reward to facilitate the implementation. TSPORec not only enhances recommendation performance but also improves computational efficiency. Extensive experiments across two models and datasets demonstrate the superb performance (up to 31.25%) and efficiency (up to 63.4%) of our approach compared with six baseline approaches. Code is available at https://github.com/WNQzhu/TSPORec.git.
