PaperlayerView exact revision on arXiv

Source-linked AI summary

Representation Learning with Large Language Models for Recommendation

Xubin Ren, Wei Wei, Lianghao Xia, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, Chao Huang

arXiv:2310.15950v5cs.IRcs.AI

TL;DR

Existing recommenders can overlook textual information and suffer from noisy implicit feedback, while direct LLM reranking faces practical limitations. RLMRec addresses this with LLM-generated profiles and cross-view alignment between collaborative and semantic representations, with theoretical support and reported improvements across integrated recommenders.

  • Problem

    ID-based graph recommenders may overlook textual information, while noisy implicit feedback and direct LLM integration create representation-learning and practical deployment challenges.

  • Method

    RLMRec generates user and item profiles with LLMs and aligns CF-side relational embeddings with LLM-side semantic representations using contrastive or generative mutual-information-based objectives.

  • Results

    RLMRec consistently improves performance when integrated with backbone recommenders, while direct ChatGPT reranking performs worse than original LightGCN results.

  • Takeaways & Limitations

    LLM knowledge is used to enhance existing recommender representations rather than directly replacing the recommender with LLM reranking.

  • Takeaways & Limitations

    The theoretical framework identifies obtaining effective user/item descriptions and modeling the density ratio as two challenges.

Abstract

from arXiv · show

Recommender systems have seen significant advancements with the influence of deep learning and graph neural networks, particularly in capturing complex user-item relationships. However, these graph-based recommenders heavily depend on ID-based data, potentially disregarding valuable textual information associated with users and items, resulting in less informative learned representations. Moreover, the utilization of implicit feedback data introduces potential noise and bias, posing challenges for the effectiveness of user preference learning. While the integration of large language models (LLMs) into traditional ID-based recommenders has gained attention, challenges such as scalability issues, limitations in text-only reliance, and prompt input constraints need to be addressed for effective implementation in practical recommender systems. To address these challenges, we propose a model-agnostic framework RLMRec that aims to enhance existing recommenders with LLM-empowered representation learning. It proposes a recommendation paradigm that integrates representation learning with LLMs to capture intricate semantic aspects of user behaviors and preferences. RLMRec incorporates auxiliary textual signals, develops a user/item profiling paradigm empowered by LLMs, and aligns the semantic space of LLMs with the representation space of collaborative relational signals through a cross-view alignment framework. This work further establish a theoretical foundation demonstrating that incorporating textual signals through mutual information maximization enhances the quality of representations. In our evaluation, we integrate RLMRec with state-of-the-art recommender models, while also analyzing its efficiency and robustness to noise data. Our implementation codes are available at https://github.com/HKUDS/RLMRec.

1 INTRODUCTION

Graph-based recommenders capture complex user-item relationships but rely heavily on IDs and implicit feedback, limiting the use of textual information and introducing noise. RLMRec addresses these issues by using LLMs for profiling and aligning semantic and collaborative representations.

  • Motivation: Graph-based recommenders use mapped user/item IDs and binary interaction matrices, potentially overlooking valuable textual information.This can reduce the informativeness of learned representations.
  • Motivation: Implicit feedback may contain noise and bias, complicating user-preference learning.
  • Motivation: Prompt-based LLM recommendation methods face efficiency and precision limitations, alongside scalability and input-constraint challenges.
  • Motivation: On Amazon reranking, ChatGPT refined recommendations performed worse than the original LightGCN results.LightGCN retrieved 50 candidates, which ChatGPT reranked for Top-10 and Top-20 recommendation.
  • Proposed direction: RLMRec uses representation learning as a bridge between LLM-based semantic information and existing recommender systems.Its framework incorporates textual signals, LLM-powered profiling, and cross-view alignment.

2 RELATED WORK

GNN-enhanced collaborative filtering models use historical user-item interactions as bipartite graphs to capture high-order collaborative relationships. However, sparse and noisy implicit feedback challenges these graph-based methods.

  • GNN-enhanced Collaborative Filtering: GNN-enhanced collaborative filtering constructs bipartite graphs from historical user-item interactions to capture high-order collaborative relationships.Examples include NGCF, GCCF, and LightGCN.
  • GNN-enhanced Collaborative Filtering: Sparsity and noise in implicit feedback data pose challenges for graph-based recommendation methods.

3 METHODOLOGY

RLMRec models noisy collaborative interactions together with LLM-derived semantic profiles, then aligns the two representation views through mutual-information-based objectives. Its profiling process uses textual attributes, reviews, and collaborative information to generate user and item descriptions.

  • Collaborative Filtering: Collaborative filtering learns user and item representations from observed interactions using a recommender model.Each user and item receives an initial embedding before representation learning.
  • Theoretical Basis: Noisy interactions can distort learned representations, motivating auxiliary textual cues and a hidden prior belief beneficial for recommendation.
  • Theoretical Basis: Theorem 1 equates maximizing posterior probability with maximizing mutual information between CF-side representation e and LLM-side semantic representation s.
  • Cross-view Alignment: RLMRec optimizes corresponding collaborative and textual representation pairs using mutual-information objectives and density-ratio modeling.The framework considers contrastive and generative approaches for this alignment.
  • Profile Generation: User and item profiles are designed to capture interaction preferences and item characteristics through LLM processing of textual and collaborative information.Item profiles use descriptions, attributes, and reviews; user profiles use sampled interacted items and user reviews.
  • Profile Generation: Item-profile generation combines available descriptions, attributes, and reviews, substituting sampled reviews when descriptions are missing.
  • Profile Generation: User-profile generation samples interacted items and organizes their textual attributes and user reviews into an LLM prompt.The construction is intended to provide insight into users’ genuine preferences.

3.3 Density Ratio Modeling for Mutual Information Maximization

RLMRec models mutual-information density ratios to align LLM-derived semantic representations with collaborative-filtering representations. It combines contrastive and generative alignment, then integrates these objectives with existing recommender losses in a model-agnostic framework.

  • Density Ratio Modeling: User/item profiles are encoded into semantic representations using a text embedding model that produces fixed-length vectors preserving textual meaning and context.The profiles are designed to capture interaction preferences before encoding.
  • Density Ratio Modeling: The density ratio f(s_i, e_i) serves as a positive real-valued similarity score between semantic and collaborative representations, supporting mutual-information maximization.More accurate density-ratio modeling is intended to improve alignment and mitigate noisy signals.
  • Contrastive Alignment: Contrastive alignment maps semantic representations into the collaborative feature space, pulls positive pairs together, and treats other batch samples as negatives.The mapping uses a multilayer perceptron and cosine similarity.
  • Generative Alignment: Generative alignment uses masked embeddings and one-direction reconstruction to recover semantic representations for randomly masked users or items.This approach models the density ratio through a masked-autoencoder-inspired reconstruction process.
  • Alignment Framework: The framework combines contrastive and generative alignment to connect LLM knowledge with user-preference understanding through collaborative relational signals and text-based behavior semantics.The two variants are named RLMRec-Con and RLMRec-Gen and are evaluated across multiple tasks.
  • Model-agnostic Learning: Because it optimizes collaborative and semantic representations independently of a recommender’s architecture, RLMRec can enhance existing collaborative-filtering models by adding its alignment objective to the recommender loss.Minimizing the overall objective corresponds to maximizing the stated mutual information.

4 EVALUATION

RLMRec is evaluated across multiple datasets, backbone recommenders, tasks, noise levels, and efficiency settings. The results generally show improved recommendation performance, robustness to noisy interactions, global relationship capture, and manageable training overhead.

  • Experimental Settings: Evaluations use three public datasets, all-rank assessment, Recall@N and NDCG@N, and six state-of-the-art collaborative filtering backbones.The experiments include Amazon-book, Yelp, and Steam, with fixed training settings and model-specific hyperparameter search.
  • Performance Comparison (RQ1): RLMRec consistently improves the performance of backbone recommenders, with contrastive alignment often stronger for GCCF and SimGCL and generative alignment more effective with AutoCF.The authors attribute gains to LLM-based user/item profiling and cross-view mutual-information maximization.
  • Performance Comparison (RQ1): RLMRec is compared with KAR using shared semantic representations and LightGCN and SGL backbones.KAR is described as an LLM-enhanced user-behavior modeling approach for CTR tasks.
  • Ablation Study (RQ2): Shuffling semantic representations tests whether alignment between semantic and collaborative signals contributes to performance, across multiple embedding models and four backbones.The study uses text-embedding-ada-002, Contriever, and Instructor with LightGCN, GCCF, SimGCL, and DCCF.
  • Performance w.r.t. Noisy Data (RQ3): Both RLMRec-Con and RLMRec-Gen outperform LightGCN at every tested noise level from 5% to 25%, with RLMRec-Con showing greater resistance to noise.The generative method may introduce additional noise through node masking.
  • Analysis of Training Efficiency (RQ5): RLMRec adds about 10% to 20% time over the original model for larger recommenders, while RLMRec-Gen is consistently faster than RLMRec-Con.The difference is attributed to batch-size-dependent contrastive computation versus masking fewer nodes in the generative approach.

5 CONCLUSION

The conclusion presents RLMRec as a model-agnostic framework that combines LLM-derived semantics with collaborative representations. It uses contrastive and generative alignment to reduce feature noise and is evaluated on real-world datasets.

  • 5 CONCLUSION: RLMRec combines collaborative profile generation, reasoning-driven prompts, and contrastive or generative alignment between CF and LLM representations.The framework is designed to improve recommender-system representation learning while reducing feature noise.

A SUPPLEMENTARY MATERIAL

The supplementary material provides implementation details for training, prompt design, profile generation, and the reranking analysis.

  • A SUPPLEMENTARY MATERIAL: The supplementary material includes pseudocode, prompt examples, item-to-user profile-generation details, and reranking-task experiments.These materials document both the framework implementation and the earlier reranking analysis.

A.1 Pseudocode of RLMRec

The pseudocode section describes training procedures for RLMRec-Con and RLMRec-Gen after semantic embeddings are generated from user and item profiles. The two variants differ in how they compute mutual-information objectives.

  • A.1 Pseudocode of RLMRec: RLMRec-Con and RLMRec-Gen preprocess profiles, generate semantic embeddings, and then follow separate training algorithms.Algorithm 1 covers RLMRec-Con, while Algorithm 2 covers RLMRec-Gen.
  • A.1 Pseudocode of RLMRec: RLMRec-Gen randomly masks users or items before CF encoding, whereas RLMRec-Con computes its mutual-information objective contrastively over all batch users and items.The generative objective uses masked representations; the contrastive approach models a density ratio for positives and negatives.

A.2 Details of Profile Generation

The profile-generation process uses LLMs to create user and item profiles from dataset information. Although illustrated with Amazon-book examples, the overall generation process is intended to remain consistent.

  • LLMs generate both user and item profiles through a general interaction paradigm.The system prompt instructs the model for the profile-generation task.
  • Amazon-book examples illustrate the profile-generation process for users and items.
  • The generation process is designed to remain consistent beyond the showcased dataset.

A.2.1 Example of the Generated Item Profile.

Item profiles summarize which users may be interested in a book, while user profiles use item preferences and user feedback to characterize interests.

  • Item profile generation: Item profiles summarize the types of users likely to appreciate each book.The item-generation instruction is shared across items.
  • Item profile generation: Item-profile generation uses each book’s title and original dataset description as input.
  • User profile generation: User profiles follow an item-to-user paradigm that leverages previously generated item profiles.
  • User profile generation: User-profile prompts combine users’ feedback on items with the items’ generated profiles.

A.3 Analysis on the Reranking Task with LLMs

The LLM reranking analysis uses prompts to select recommendations from LightGCN-retrieved candidates, but exposes hallucination, limited textual preference information, and scalability constraints.

  • Reranking setup: The reranking setup asks an LLM to recommend 20 books from a 40-item candidate set retrieved by LightGCN.The prompt provides previously interacted items together with indexed candidate items.
  • Reranking setup: Candidate items are represented with integer indices and textual information such as titles to support accurate selection.The JSON output requires a list of item indices rather than titles.
  • Observed problems: LLM outputs can include non-existent items despite indexed candidate lists, undermining reranking quality.
  • Observed problems: The LLM produces fewer correctly recommended items than LightGCN because available textual information provides less preference signal than collaborative information.
  • Limitations: Raw textual augmentation is constrained by token limits, missing or noisy descriptions, and increased computational cost.
Loading 2310.15950v5…