Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Reindex-Then-Adapt: Improving Large Language Models for Conversational Recommendation

Zhankui He, Zhouhang Xie, Harald Steck, Dawen Liang, Rahul Jha, Nathan Kallus, Julian McAuley

TL;DR

This work tackles distribution misalignment in conversational recommender systems using large language models by introducing Reindex-Then-Adapt (RTA). RTA first reindexes multi-token item titles into single-token embeddings, enabling efficient logit computation, and then adapts the resulting distributions through bias-term adjustments or gating with traditional RecSys models to align with target data distributions. By treating LLMs as Differentiable Search Indexes and leveraging both L2I/L2R content knowledge and conventional recommendations, the approach achieves substantial accuracy gains across INSPIRED, ReDIAL, and Reddit-V1.5 datasets with favorable efficiency and flexibility characteristics. The results show that simple, dataset-aware adaptation strategies can meaningfully improve CRS performance, demonstrating practical impact in aligning LLM-driven recommendations with dynamic platform distributions and enabling controllability and fairness in recommendations.

Abstract

Large language models (LLMs) are revolutionizing conversational recommender systems by adeptly indexing item content, understanding complex conversational contexts, and generating relevant item titles. However, controlling the distribution of recommended items remains a challenge. This leads to suboptimal performance due to the failure to capture rapidly changing data distributions, such as item popularity, on targeted conversational recommendation platforms. In conversational recommendation, LLMs recommend items by generating the titles (as multiple tokens) autoregressively, making it difficult to obtain and control the recommendations over all items. Thus, we propose a Reindex-Then-Adapt (RTA) framework, which converts multi-token item titles into single tokens within LLMs, and then adjusts the probability distributions over these single-token item titles accordingly. The RTA framework marries the benefits of both LLMs and traditional recommender systems (RecSys): understanding complex queries as LLMs do; while efficiently controlling the recommended item distributions in conversational recommendations as traditional RecSys do. Our framework demonstrates improved accuracy metrics across three different conversational recommendation datasets and two adaptation settings

Reindex-Then-Adapt: Improving Large Language Models for Conversational Recommendation

TL;DR

This work tackles distribution misalignment in conversational recommender systems using large language models by introducing Reindex-Then-Adapt (RTA). RTA first reindexes multi-token item titles into single-token embeddings, enabling efficient logit computation, and then adapts the resulting distributions through bias-term adjustments or gating with traditional RecSys models to align with target data distributions. By treating LLMs as Differentiable Search Indexes and leveraging both L2I/L2R content knowledge and conventional recommendations, the approach achieves substantial accuracy gains across INSPIRED, ReDIAL, and Reddit-V1.5 datasets with favorable efficiency and flexibility characteristics. The results show that simple, dataset-aware adaptation strategies can meaningfully improve CRS performance, demonstrating practical impact in aligning LLM-driven recommendations with dynamic platform distributions and enabling controllability and fairness in recommendations.

Abstract

Large language models (LLMs) are revolutionizing conversational recommender systems by adeptly indexing item content, understanding complex conversational contexts, and generating relevant item titles. However, controlling the distribution of recommended items remains a challenge. This leads to suboptimal performance due to the failure to capture rapidly changing data distributions, such as item popularity, on targeted conversational recommendation platforms. In conversational recommendation, LLMs recommend items by generating the titles (as multiple tokens) autoregressively, making it difficult to obtain and control the recommendations over all items. Thus, we propose a Reindex-Then-Adapt (RTA) framework, which converts multi-token item titles into single tokens within LLMs, and then adjusts the probability distributions over these single-token item titles accordingly. The RTA framework marries the benefits of both LLMs and traditional recommender systems (RecSys): understanding complex queries as LLMs do; while efficiently controlling the recommended item distributions in conversational recommendations as traditional RecSys do. Our framework demonstrates improved accuracy metrics across three different conversational recommendation datasets and two adaptation settings
Paper Structure (50 sections, 6 equations, 6 figures, 5 tables)

This paper contains 50 sections, 6 equations, 6 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Task Formulation
  4. Differentiable Search Index (DSI)
  5. Item Indexing: LLMs Show Sufficient Item Content Knowledge
  6. Observation
  7. Impact
  8. Item Recommendation: LLMs Show Severe Distribution Misalignment
  9. Observation
  10. Impact
  11. Framework
  12. Overview
  13. Reindex Step: Single-Token Items in LLMs
  14. Identify Item Indices.
  15. Aggregate Multi-Token Embeddings
  16. ...and 35 more sections

Figures (6)

  • Figure 1: Representative items (The Dark Knight and Black Panther) demonstrate popularity misalignments between the dataset (ReDIAL li2018redial) and the LLM (Llama2-7b touvron2023llama). This misalignment implies a significant room for recommendation accuracy improvement. Our Reindex-Then-Adapt (RTA) framework addresses this gap by aligning the distributions (e.g., item popularites), leading to substantial accuracy improvements.
  • Figure 2: Item Indexing tasks by using movie descriptions from WikiPedia auer2007dbpedia to prompt movie titles. We tested MPT-7b, Mistrial-7b , Llama2-7b and GPT-Turbo-3.5 models and group the accuracy by the range of occurrences of the movies in ReDIAL CRS dataset li2018redial. We measure the performance with HIT@5, i.e., whether the target movie in the top-K movie list generated by the LLMs to reflect the movie knowledge stored in the parameters of the LLMs.
  • Figure 3: Visualization of item monthly relative popularity from Reddit-Movie he2023large datasets, since this dataset is the only CRS dataset with long-range timestamps in the wild. Item popularities are shown changing overtime rapidly.
  • Figure 4: Reindex-Then-Adapt (RTA) Framework. LLMs can generate a list of item titles as the recommendations given the conversation contexts. To further improve the accuracy and controllability, we conduct (1) reindex step: reindexing the item (e.g., movie) titles in LLMs as single tokens to obtain the predicted logit vectors efficiently; (2) adapt step: adapting the recommenders towards target data distributions effectively with multiple options on the logit vectors such as adjusting bias terms or combining RecSys models with Gating mechanism.
  • Figure 5: Different methods to represent items in LLMs with single-token embeddings and the related recommendation accuracy HIT@5 after the reindex step.
  • ...and 1 more figures