基于负反馈修正的多轮对话推荐系统

Multi-Round Conversational Recommendation System Based on Negative Feedback Correction

传统的推荐系统从交互历史中挖掘用户兴趣,面临着无法动态地获取用户实时偏好和细粒度偏好的问题,近年对话推荐系统领域的兴起为此问题提供了新的解决方案.对话推荐系统优势在于其可以动态地和用户进行交互,并在交互过程中获取用户的实时偏好,从而提高推荐系统准确率,提升用户体验.然而对话推荐系统相关研究工作中缺乏对负反馈的充分利用,难以对用户偏好表示进行细粒度的修正,即难以有效平衡用户长期偏好和实时偏好之间的关系,同时存在属性候选集过大导致交互轮次过多的问题.因此,本文基于经典的对话推荐框架CPR(Conversational Path Reasoning)提出了一种能够有效利用用户负反馈的对话推荐模型NCPR(Negative-feedback-guide Conversational Path Reasoning).不同于现有的对话推荐系统工作,NCPR能够充分利用用户在交互过程中给出的属性粒度和物品粒度的负反馈对用户的偏好表示进行动态的修正.此外,CPR将对话推荐建模为一个图上的路径推理问题,NCPR使用协同过滤算法基于属性粒度的负反馈对属性候选集进行重排序,在利用图结构的自然优势限制属性候选集大小的同时,进一步减少候选属性空间大小.四个基准数据集上的实验结果表明,NCPR在推荐准确率和平均交互轮次两个评价指标上的表现优于先进的基线模型.最后,我们设计并实现了一个网页端的对话推荐系统,与在线用户进行交互产生推荐结果,证明了NCPR在真实的对话推荐场景下的有效性.

Traditional recommendation systems can only estimate user preferences and model user preferences on items from past interaction history,thus suffering from the limitations of obtaining real-time and fine-grained user preferences.Conversational recommendation systems(CRS),which introduce conversational technology into recommendation systems,provide a new solution to this problem in recent years.Unlike traditional recommendation systems which,due to their static way of working,cannot answer what the user’s current preferences are and what the user’s reasons are for buying an item,CRS are interactive recommendation systems,which have the advantage of dynamically interacting with users and obtaining their real-time preferences in the process of interaction,understanding what they currently like and what their reasons are for liking an item,thus allowing the recommendation system to quickly understand the user’s intent and make recommendations that the user desires,enhancing the user’s experience and trust in the system.However,current CRS studies more often uses positive feedback on the granularity of attrib utes given by the user to model the user’s real-time preferences,ignoring the impact of negative user feedback on modeling the user’s real-time preferences,but negative feedback is an important part of user feedback and can also indicate real-time user preferences,making it difficult to make fine-grained corrections to the user preference representation,which means it is difficult to effectively balance the relationship between users’long-term preference and real-time preference.At the same time,current state-of-the-art work in the field only takes advantage of the natural advantages of graph structure to limit the size of the attribute candidate set,which suffers from the problem of too many interaction rounds due to the large attribute candidate set.To address the problems mentioned above,we propose a conversational recommendation model NCPR based on the classical conversational recommendation framework CPR,which can make full use of all feedback given by the user during the interaction to correct the user’s preference representation,including positive feedback at the attribute granularity,negative feedback at the attribute granularity,and negative feedback at the item granularity.In addition,CPR models conversation recommendation as a path inference problem on a graph,i.e.CRS can only select attribute nodes adjacent to the current node to ask the user.This approach helps CRS to limit the size of the attribute candidate set in a single round of decision making.NCPR uses an attribute-based collaborative filtering algorithm to reorder the attribute candidate set based on negative feedback from the attribute granularity,i.e.removing those attributes in the attribute candidate set that are similar to those rejected by the user,which can further reduce the candidate attribute space size while taking advantage of the natural advantages of graph structure to limit the size of the attribute candidate set.Experimental results on four benchmark CRS datasets show that the proposed method significantly outperforms state-of-the-art baselines in terms of both evaluation metrics.In addition,we design and implement a web-side conversational recommendation system that interacts with online users to generate recommendation results,demonstrating the effectiveness of NCPR in a real-world conversational recommendation scenario.