ProEx: A Unified Framework Leveraging Large Language Model with Profile Extrapolation for Recommendation
Yi Zhang, Yiwen Zhang, Yu Wang, Tong Chen, Hongzhi Yin
TL;DR
ProEx addresses the instability and bias of single LLM-generated profiles in recommendation by generating multiple diverse profiles per user/item via chain-of-thought reasoning and extrapolating in language space. It introduces environments that linearly mix profiles to extract invariant preferences, and employs cross-space alignment plus contrastive regularization to fuse language-space profiles with traditional collaborative representations. Across discriminative and generative baselines on three datasets, ProEx consistently improves performance over strong LLM-enhanced and standard methods, with ablations highlighting the value of profile extrapolation, environment-based mixing, and CoT-guided generation. The framework provides a practical, model-agnostic path to leverage rich semantic priors from LLMs while mitigating bias and instability, with guidance on hyperparameters and representation choices.
Abstract
The powerful text understanding and generation capabilities of large language models (LLMs) have brought new vitality to general recommendation with implicit feedback. One possible strategy involves generating a unique user (or item) profile from historical interaction data, which is then mapped to a semantic representation in the language space. However, a single-instance profile may be insufficient to comprehensively capture the complex intentions behind a user's interacted items. Moreover, due to the inherent instability of LLMs, a biased or misinterpreted profile could even undermine the original recommendation performance. Consequently, an intuitive solution is to generate multiple profiles for each user (or item), each reflecting a distinct aspect of their characteristics. In light of this, we propose a unified recommendation framework with multi-faceted profile extrapolation (ProEx) in this paper. By leveraging chain-of-thought reasoning, we construct multiple distinct profiles for each user and item. These new profiles are subsequently mapped into semantic vectors, extrapolating from the position of the original profile to explore a broader region of the language space. Subsequently, we introduce the concept of environments, where each environment represents a possible linear combination of all profiles. The differences across environments are minimized to reveal the inherent invariance of user preferences. We apply ProEx to three discriminative methods and three generative methods, and conduct extensive experiments on three datasets. The experimental results demonstrate that ProEx significantly enhances the performance of these base recommendation models.