Prompting LLMs for Zero-Shot Next-Item Recommendation
Loading...
Date
2023-12-01
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Recommendation systems can offer personalized suggestions based on user data, enhancing user experiences. Sequential recommendation is a specific subcategory of recommendation systems that focuses on the order and context of previous interactions to offer a sequence of item suggestions. Large language models (LLMs) have demonstrated exceptional prowess in tasks involving commonsense reasoning, knowledge utilization, and task generalization. In particular, they have exhibited remarkable zero-shot performance in numerous natural language processing (NLP) tasks, showcasing their capacity to recommend without extensive training data. This thesis explores the untapped potential of integrating LLMs into sequential recommendations to enhance performance and elevate the user experience. However, this endeavor faces three primary challenges: (i) The huge recommendation space poses significant challenges to LLM-based recommendations, (ii) LLMs face a fundamental constraint that it is not feasible to include all possible items in the prompt, and (iii) for items that appear infrequently in the training of an LLM, it can be challenging to model these items well to make recommendations. In this thesis, we propose a prompting strategy called zero-shot next-item recommendation (NIR) prompting to guide LLM to make next-item recommendations. Specifically, NIR-based strategies involve using external modules to generate candidate projects based on user filtering. Our strategy uses a 4-step prompt to guide GPT-3 to learn the user’s preferences through the user’s past interaction history and recommend a ranked list of K movies. We evaluate the proposed method using GPT-3 on the MovieLens 100K dataset and show that it achieves strong zero-shot performance. Additionally, we also demonstrate that LLMs can be affected by biases like position bias and popularity bias. By employing specialized prompting and bootstrapping strategies, these biases can be mitigated.
Description
Keywords
LLMs, Recommendation System, Sequential Recommendation