Harnessing Generative Large Language Models for Dynamic Intention Understanding in Recommender Systems: Insights From a Client–Designer Interaction Case Study

Abstract

Generative large language models (GLLMs) have achieved extreme success in the academic community of recommender systems. However, the application of such a powerful tool in the industrial world is still nascent. In Chinese home renovation industry, advisory consultants engage in offline conversations to fully understand the intentions of potential clients before subsequently recommending designers to them. Although conventional recommender systems can somewhat substitute for the consultants, they fall short in addressing two significant challenges. First, clients frequently revise their intentions during conversations, complicating the accurate capture of key intentions. Second, the process of recommending designers, which relies heavily on consultants’ manual efforts, is not only time-consuming but also prone to inaccuracies. To address the challenges, we present a recommendation agent, named DCICDRec, which leverages the robust conversational understanding and generation capabilities of the large language model MOSS. The creation of this agent involves two key steps. The first step is to prepare the corpus from the renovation domain by organizing it into conversational graphs, to which balanced sampling and profile normalization mechanisms are applied. This preparation ensures that the corpus is well-structured and unbiased before proceeding to fine-tune MOSS. The second step is to utilize the fine-tuned MOSS as a recommendation agent. In this capacity, the agent engages in conversations with potential clients and recommends designers, providing detailed reasons for each recommendation. Furthermore, if the client is dissatisfied with the recommended designers, the agent will delve deeper into understanding the client's true intentions and continually update the recommendations until the client is satisfied. We evaluate the agent's effectiveness on a real dialog dataset CRM between clients and consultants, as well as two publicly available datasets, INSPIRED and ReDIAL. Through comprehensive experiments with six baseline models, the DCICDRec agent demonstrate superior performances on the three datasets. Such experimental achievements indicate that the DCICDRec agent holds significant potential for generalization and commercial value. Moreover, the results of case study with 11 offline tests illustrate the scalability and efficiency of the agent in real-time scenarios.