| 摘要: |
| 推荐系统的目标是从物品数据库中，选择出与用户兴趣偏好相匹配的子集，缓解用户面临的“信息过载”问题。因而近年来推荐系统越来越多地应用到电商、社交等领域，展现出巨大的商业潜力。传统推荐系统中，系统对用户的认知往往来源于历史交互记录，例如点击率或者购买记录，这是一种隐式用户反馈。对话推荐系统能够通过自然语言与用户进行多轮对话，逐步深入挖掘其兴趣偏好，从而向对方提供高质量的推荐结果。相比于传统推荐系统，对话推荐系统主要有两方面的不同。其一，对话推荐系统能够利用自然语言与用户进行语义上连贯的多轮对话，提升了人机交互中的用户体验；其二，系统能够询问特定的问题直接获取用户的显式反馈，从而更深入地理解用户兴趣偏好，提供更可靠的推荐结果。目前已经有不少工作在不同的问题设定下对该领域进行了探索，然而尽管如此，这些工作仍仅局限于关注当前正在进行的对话，忽视了过去交互记录中蕴涵的丰富信息，导致对用户偏好建模的不充分。为了解决这个问题，本文提出了一个面向用户偏好建模的个性化对话推荐算法框架，通过双线性模型注意力机制与自注意力层次化编码结构进行用户偏好建模，从而完成对候选物品的排序与推荐。本文设计的模型结构能够在充分利用用户历史对话信息的同时，权衡历史对话与当前对话两类数据的重要性。丰富的用户相关信息来源使得推荐结果在契合用户个性化偏好的同时，更具备多样性，从而缓解“信息茧房”等现象带来的不良影响。基于公开数据集的实验表明了本文方法在个性化对话推荐任务上的有效性。 |
| 关键词: 对话推荐系统 用户建模 图神经网络 注意力机制 |
| DOI：10.19363/J.cnki.cn10-1380/tn.2021.09.05 |
| 投稿时间：2021-05-06修订日期：2021-08-09 |
| 基金项目:本课题得到国家自然科学基金（No.61872369，No.61832017）资助。 |
|
| Modeling User Preference for Personalized Conversational Recommendation |
| SHANG Chenzhan,ZHAO Xin |
| School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan 430074, China;Gaoling School of Artificial Intelligence, Renmin University of China, Beijing 100872, China |
| Abstract: |
| Recommender Systems aim to select the subset that matches users' preference from the item database, so as to alleviate the problem of "information overload" faced by users. Therefore, in recent years, recommender systems have been increasingly applied to e-commerce, social networking and other fields, showing great commercial potential. In conventional recommender systems, the information of users is often derived from the historical interaction records of items, such as click rates and purchase history, which can be seen as a kind of delayed and implicit feedback from users. Conversational recommender systems (CRS) are able to gradually dig into users' preference through multi-turn dialogues in natural language, so as to recommend high-quality items to them. Compared to conventional recommender systems, CRS exhibit major differences in two aspects. Firstly, CRS are able to interact for multiple semantically coherent rounds with users in natural language, which improves user experience during human-computer interaction. Secondly, it becomes possible for CRS to obtain explicit feedback from users by asking specific questions, which helps system to have a deeper understanding of users' interest and preference, and provide more reliable recommendation results to them. Many works have explored conversational recommendation under different problem settings. However, these works only focus on the ongoing dialogue, ignoring the rich information contained in the past conversations, which leads to insufficiency in user modeling. To address these issues, we propose an end-to-end framework named PCR, which stands for user preference modeling based Personalized Conversational Recommender. PCR is able to model users through the context-attentive layer with bilinear scoring function and the self-attentive hierarchical encoding structure. Our proposed model can make full use of historical dialogues while weighing the importance of the historical dialogues and the current dialogue. More information sources about users make the recommendation results more diversified while conforming to users' preferences, which will alleviate the impact of "information cocoon". Extensive experiments have demonstrated the effectiveness of our approach on personalized conversational recommendation task. |
| Key words: conversational recommender system user modeling graph neural network attention mechanism |
