Traditional recommender systems have predominantly relied on identity representations (IDs) to characterize users and items. In contrast, the emergence of pre-trained language model (PLM) en-coders has significantly enriched the modeling of contextual item descriptions. While PLMs excel in addressing few-shot, zero-shot, and unified modeling scenarios, they often overlook the critical collaborative filtering signal. This omission gives rise to two pivotal challenges: (1) Collaborative Contextualization, aiming for the seamless integration of collaborative signals with contextual representations. (2) The necessity to bridge the representation gap between ID-based and contextual representations while preserving their contextual semantics. In this paper, we introduce CollabContext, a novel model that skillfully merges collaborative filtering signals with contextual representations, aligning these representations within the contextual space while retaining essential contextual semantics. Experimental results across three real-world datasets showcase substantial improvements. Through its capability in collaborative contextualization, CollabContext demonstrates remarkable enhancements in recommendation performance, particularly in cold-start scenarios. The code is available after the conference accepts the paper.

翻译：传统推荐系统主要依赖身份标识（ID）来表征用户和项目。相比之下，预训练语言模型（PLM）编码器的出现显著丰富了上下文项目描述的建模能力。尽管PLM在少样本、零样本及统一建模场景中表现优异，但它们通常忽略了关键的协同过滤信号。这一缺失引发了两个关键挑战：(1) 协同上下文化，旨在实现协同信号与上下文表示的无缝整合；(2) 有必要弥合基于ID的表示和上下文表示之间的表征鸿沟，同时保留其上下文语义。在本文中，我们提出了CollabContext——一种新颖模型，它巧妙地将协同过滤信号与上下文表示融合，在保持核心上下文语义的同时，将这些表示对齐到上下文空间。在三个真实世界数据集上的实验结果表明，该方法带来了显著提升。通过其在协同上下文化方面的能力，CollabContext在推荐性能上展现出了显著改进，尤其是在冷启动场景中。代码将在会议接收论文后公开。