Abstract
ProEx is a unified framework designed to enhance recommender systems by leveraging Large Language Models (LLMs) through a process called profile extrapolation. While LLMs can summarize user interactions into profiles, single-instance profiling is often biased or inconsistent. ProEx addresses this by generating multiple diverse profiles via a chain-of-thought procedure and extrapolating them in the semantic space. By applying environment-based invariant learning, the model identifies robust user preferences that persist across different profile facets. Experimental results on major benchmarks demonstrate that ProEx significantly outperforms state-of-the-art baselines in both recommendation accuracy and training efficiency.
Introduction
Modern recommender systems often struggle with "data sparsity," where a lack of user interactions makes it difficult to predict preferences. Recently, Large Language Models (LLMs) have been used to bridge this gap by summarizing user histories into textual profiles. However, research indicates that relying on a single LLM-generated profile is inherently brittle and prone to error.
The Pitfalls of One-Shot Profiling
LLMs are susceptible to hallucinations, popularity bias, and semantic inconsistency. If a single profile misinterprets a user's intent or focuses on a niche interest, that error is propagated into the recommendation model. As illustrated in the following figure, a single LLM response can be "meaningless" or overly restrictive, failing to capture the multi-faceted nature of human behavior.
The ProEx Framework
The ProEx (Profile Extrapolation) framework addresses these issues by generating and integrating multiple diverse profiles for every user and item. The process consists of three main stages:
1. Multi-faceted Profile Construction
Instead of a single summary, ProEx uses a four-step Chain-of-Thought (CoT) procedure. It generates an original profile, extracts key positive/negative aspects, analyzes latent preferences, and finally re-generates several new profiles that are lexically distinct. This process uncovers latent preferences, such as "emotional healing" or "human connection," that might be missing from a simple summary.
2. Profile Extrapolation
These text profiles are converted into semantic vectors. To ensure diversity, ProEx uses a contrastive regularization loss that "pushes" the profiles apart in the latent space. This "extrapolates" the user's preference region, covering a broader range of potential interests and exploring the "convex hull" of a user's possible interests.
3. Environment-based Invariant Learning
To identify the core user preference amidst these diverse profiles, ProEx introduces "environments"—different mixtures of the generated profiles. By minimizing performance variance across these mixtures, the model learns to focus on invariant features that remain consistent regardless of which specific profile facet is emphasized.
Experimental Evaluation
The authors evaluated ProEx on three public benchmarks: Amazon-Book, Yelp, and Steam. ProEx is model-agnostic and was successfully applied to both discriminative models (like LightGCN) and generative models (like Mult-VAE).
Key Results
- Significant Performance Gains: When added to LightGCN, ProEx improved Recall@20 by 8.67% on Amazon-Book, 13.05% on Yelp, and 9.11% on Steam.
- Superiority over Baselines: ProEx outperformed recent LLM-based methods like LLMRec and DMRec, proving that multiple extrapolated profiles are more effective than sophisticated single-profile methods.
- Improved Efficiency: Models using ProEx converged to their peak performance faster than base models, despite the added complexity of processing multiple profiles.
Conclusion
ProEx demonstrates that the diversity of LLM outputs can be leveraged as a strength rather than a weakness. By extrapolating user profiles and focusing on invariant preferences, the framework provides a robust and high-performing solution for modern recommendation challenges, effectively mitigating the biases inherent in single-instance LLM outputs.
Summary updated