Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Bursting Filter Bubble: Enhancing Serendipity Recommendations with Aligned Large Language Models

Yunjia Xi, Muyan Weng, Wen Chen, Chao Yi, Dian Chen, Gaoyang Guo, Mao Zhang, Jian Wu, Yuning Jiang, Qingwen Liu, Yong Yu, Weinan Zhang

TL;DR

The paper tackles the filter bubble problem in recommender systems by introducing SERAL, a three-stage framework that leverages aligned large language models to deliver serendipity without sacrificing utility. It combines Cognition Profile Generation to compress long user histories, SerenGPT Alignment with Collaborative Data Intervention to align serendipity judgments to human preferences, and Nearline Adaptation to enable efficient industrial deployment via a serendipity channel and caching. Empirical results in Taobao show consistent gains in serendipity exposure, clicks, and transactions with minimal impact on overall revenue, and ablations confirm the importance of IPO-based preference alignment and CDI for data quality and diversity. The work demonstrates the practical viability of using aligned LLMs for serendipity in large-scale RSs and provides a deployable blueprint for breaking the filter bubble in industry-scale systems.

Abstract

Recommender systems (RSs) often suffer from the feedback loop phenomenon, e.g., RSs are trained on data biased by their recommendations. This leads to the filter bubble effect that reinforces homogeneous content and reduces user satisfaction. To this end, serendipity recommendations, which offer unexpected yet relevant items, are proposed. Recently, large language models (LLMs) have shown potential in serendipity prediction due to their extensive world knowledge and reasoning capabilities. However, they still face challenges in aligning serendipity judgments with human assessments, handling long user behavior sequences, and meeting the latency requirements of industrial RSs. To address these issues, we propose SERAL (Serendipity Recommendations with Aligned Large Language Models), a framework comprising three stages: (1) Cognition Profile Generation to compress user behavior into multi-level profiles; (2) SerenGPT Alignment to align serendipity judgments with human preferences using enriched training data; and (3) Nearline Adaptation to integrate SerenGPT into industrial RSs pipelines efficiently. Online experiments demonstrate that SERAL improves exposure ratio (PVR), clicks, and transactions of serendipitous items by 5.7%, 29.56%, and 27.6%, enhancing user experience without much impact on overall revenue. Now, it has been fully deployed in the "Guess What You Like" of the Taobao App homepage.

Bursting Filter Bubble: Enhancing Serendipity Recommendations with Aligned Large Language Models

TL;DR

The paper tackles the filter bubble problem in recommender systems by introducing SERAL, a three-stage framework that leverages aligned large language models to deliver serendipity without sacrificing utility. It combines Cognition Profile Generation to compress long user histories, SerenGPT Alignment with Collaborative Data Intervention to align serendipity judgments to human preferences, and Nearline Adaptation to enable efficient industrial deployment via a serendipity channel and caching. Empirical results in Taobao show consistent gains in serendipity exposure, clicks, and transactions with minimal impact on overall revenue, and ablations confirm the importance of IPO-based preference alignment and CDI for data quality and diversity. The work demonstrates the practical viability of using aligned LLMs for serendipity in large-scale RSs and provides a deployable blueprint for breaking the filter bubble in industry-scale systems.

Abstract

Recommender systems (RSs) often suffer from the feedback loop phenomenon, e.g., RSs are trained on data biased by their recommendations. This leads to the filter bubble effect that reinforces homogeneous content and reduces user satisfaction. To this end, serendipity recommendations, which offer unexpected yet relevant items, are proposed. Recently, large language models (LLMs) have shown potential in serendipity prediction due to their extensive world knowledge and reasoning capabilities. However, they still face challenges in aligning serendipity judgments with human assessments, handling long user behavior sequences, and meeting the latency requirements of industrial RSs. To address these issues, we propose SERAL (Serendipity Recommendations with Aligned Large Language Models), a framework comprising three stages: (1) Cognition Profile Generation to compress user behavior into multi-level profiles; (2) SerenGPT Alignment to align serendipity judgments with human preferences using enriched training data; and (3) Nearline Adaptation to integrate SerenGPT into industrial RSs pipelines efficiently. Online experiments demonstrate that SERAL improves exposure ratio (PVR), clicks, and transactions of serendipitous items by 5.7%, 29.56%, and 27.6%, enhancing user experience without much impact on overall revenue. Now, it has been fully deployed in the "Guess What You Like" of the Taobao App homepage.

Paper Structure

This paper contains 42 sections, 7 equations, 5 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Serendipity Recommendations
  4. LLM-based Recommendations
  5. Preliminaries
  6. Task Formulation
  7. Serendipity in industrial RSs
  8. Methodology
  9. Overview
  10. Cognition Profile Generation
  11. Static Profile
  12. Short-term Profile
  13. Long-term Profile
  14. SerenGPT Alignment
  15. Serendipity Data Generation
  16. ...and 27 more sections

Figures (5)

  • Figure 1: The comparison between (a) traditional recommender system and (b) our serendipity recommender system.
  • Figure 2: The overall framework of SERAL, with Cognitive Profile Generation to compresses long behavior sequences into concise profiles, SerenGPT Alignment to align LLMs' serendipity judgments with humans, and Nearline Adaptation to integrate serenGPT into industrial recommender systems efficiently.
  • Figure 3: Long-term online impact of serendipity items.
  • Figure 4: Online Performance of different caching frequency.
  • Figure 5: Prompt for short-term profile.

Theorems & Definitions (1)

  • Definition 1