Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Improving LLM-powered Recommendations with Personalized Information

Jiahao Liu, Xueshuo Yan, Dongsheng Li, Guangping Zhang, Hansu Gu, Peng Zhang, Tun Lu, Li Shang, Ning Gu

TL;DR

This paper tackles the underutilization of LLM reasoning in recommender systems by introducing CoT-Rec, a two-stage pipeline that injects Chain-of-Thought processes—user preference analysis and item perception analysis—into LLM-powered recommendations. It combines a CRM-based retrieval stage (CRM-as-Retriever) with an LLM-based ranking stage (LLM-as-Ranker), and shows how explicit personalized information extraction and utilization improve both retrieval accuracy and ranking robustness, including reduced position bias. The approach is model-agnostic, scalable (operations in extraction are offline), and employs techniques like Encode & Map embeddings, role-playing item perception, and LoRA-tuned ranking. Experimental results on three datasets demonstrate meaningful gains across retrieval and ranking, with publicly available code, highlighting practical potential for industrial deployment.

Abstract

Due to the lack of explicit reasoning modeling, existing LLM-powered recommendations fail to leverage LLMs' reasoning capabilities effectively. In this paper, we propose a pipeline called CoT-Rec, which integrates two key Chain-of-Thought (CoT) processes -- user preference analysis and item perception analysis -- into LLM-powered recommendations, thereby enhancing the utilization of LLMs' reasoning abilities. CoT-Rec consists of two stages: (1) personalized information extraction, where user preferences and item perception are extracted, and (2) personalized information utilization, where this information is incorporated into the LLM-powered recommendation process. Experimental results demonstrate that CoT-Rec shows potential for improving LLM-powered recommendations. The implementation is publicly available at https://github.com/jhliu0807/CoT-Rec.

Improving LLM-powered Recommendations with Personalized Information

TL;DR

This paper tackles the underutilization of LLM reasoning in recommender systems by introducing CoT-Rec, a two-stage pipeline that injects Chain-of-Thought processes—user preference analysis and item perception analysis—into LLM-powered recommendations. It combines a CRM-based retrieval stage (CRM-as-Retriever) with an LLM-based ranking stage (LLM-as-Ranker), and shows how explicit personalized information extraction and utilization improve both retrieval accuracy and ranking robustness, including reduced position bias. The approach is model-agnostic, scalable (operations in extraction are offline), and employs techniques like Encode & Map embeddings, role-playing item perception, and LoRA-tuned ranking. Experimental results on three datasets demonstrate meaningful gains across retrieval and ranking, with publicly available code, highlighting practical potential for industrial deployment.

Abstract

Due to the lack of explicit reasoning modeling, existing LLM-powered recommendations fail to leverage LLMs' reasoning capabilities effectively. In this paper, we propose a pipeline called CoT-Rec, which integrates two key Chain-of-Thought (CoT) processes -- user preference analysis and item perception analysis -- into LLM-powered recommendations, thereby enhancing the utilization of LLMs' reasoning abilities. CoT-Rec consists of two stages: (1) personalized information extraction, where user preferences and item perception are extracted, and (2) personalized information utilization, where this information is incorporated into the LLM-powered recommendation process. Experimental results demonstrate that CoT-Rec shows potential for improving LLM-powered recommendations. The implementation is publicly available at https://github.com/jhliu0807/CoT-Rec.

Paper Structure

This paper contains 26 sections, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. LLM-as-RSs
  4. LLM-as-Retriever
  5. LLM-as-Ranker
  6. LLM-enhanced RSs
  7. Preliminaries
  8. Methods
  9. Personalized Information Extraction
  10. User Preference Maintenance Module
  11. Item Perception Analysis Module
  12. Scalability
  13. Personalized Information Utilization
  14. Architecture
  15. CRM-as-Retriever
  16. ...and 11 more sections

Figures (4)

  • Figure 1: Overview of CoT-Rec.
  • Figure 2: User preference maintenance module.
  • Figure 3: Item perception analysis module.
  • Figure 4: Model architecture of the personalized information utilization stage in CoT-Rec. Note that we have added user embedding as an input for the CRM and included extracted user preferences and subjective perception of items in the prompt.