CoRAL: Collaborative Retrieval-Augmented Large Language Models Improve Long-tail Recommendation

TL;DR

CoRAL tackles long-tail recommendations by closing the gap between LLM reasoning and task-specific collaborative information. It introduces collaborative retrieval-augmented prompting and an RL-trained retrieval policy (DDPG) to select a minimal-sufficient set of user-item interactions that guide the LLM. The method is warm-started with short-head data to improve data efficiency and convergence. Experiments on four Amazon datasets show consistent improvements in AUC and F1, illustrating the practical value of aligning LLMs with collaborative signals for tail items.

Abstract

The long-tail recommendation is a challenging task for traditional recommender systems, due to data sparsity and data imbalance issues. The recent development of large language models (LLMs) has shown their abilities in complex reasoning, which can help to deduce users' preferences based on very few previous interactions. However, since most LLM-based systems rely on items' semantic meaning as the sole evidence for reasoning, the collaborative information of user-item interactions is neglected, which can cause the LLM's reasoning to be misaligned with task-specific collaborative information of the dataset. To further align LLMs' reasoning to task-specific user-item interaction knowledge, we introduce collaborative retrieval-augmented LLMs, CoRAL, which directly incorporate collaborative evidence into the prompts. Based on the retrieved user-item interactions, the LLM can analyze shared and distinct preferences among users, and summarize the patterns indicating which types of users would be attracted by certain items. The retrieved collaborative evidence prompts the LLM to align its reasoning with the user-item interaction patterns in the dataset. However, since the capacity of the input prompt is limited, finding the minimally-sufficient collaborative information for recommendation tasks can be challenging. We propose to find the optimal interaction set through a sequential decision-making process and develop a retrieval policy learned through a reinforcement learning (RL) framework, CoRAL. Our experimental results show that CoRAL can significantly improve LLMs' reasoning abilities on specific recommendation tasks. Our analysis also reveals that CoRAL can more efficiently explore collaborative information through reinforcement learning.

Figures (3)

• Figure 1: By text comprehension and extracting information-rich semantic features runfeng2023lkpnr, LLMs in (a) can handle long-tail items sanner2023large, but still cannot directly leverage collaborative information. To handle long-tail items in collaborative filtering-based recommender systems, by collaborative prompting, LLMs in (b) can reason the fact that even if the current item shares the same theme with previously liked items, users with similar interests still dislike this item, which provides the rationale to not recommending it.
• Figure 2: CoRAL's (DFM and WDL) learning curves on Gift Cards and Prime Pantry datasets.
• Figure 3: CoRAL's performance (AUC and F1) w.r.t number of iterations of user-item retrieval on four Amazon Product datasets.