融合多源信息的可解释性推荐模型及算法研究

With the exponential increasing of data, recommendation system is nowadays ubiquitous in various domains because it has ability to provide users with the information of potential interest. Most existing recommendation systems attempt to improve the prediction accuracy but ignore the generation of reasonable explanations. Recently, researchers have turned their attention to explaining recommendations, and help the users make informed decisions easily. In order to increase conversion rates and lead to more satisfied users, it is urgent and important to design recommendation systems with both high prediction accuracy and strong explainable recommendation. Based on the theories including statistical learning, cooperation learning, graph computing, optimization and etc., in this project, we will analyze the main properties of recommendation data: multi-source, large-scale, dynamic and etc. and do a series of researches. Firstly, we will propose the explainable recommendation models by integrating sentiment-based topic extraction and latent user/item factors learning, study the correlations between topics and latent factors, and provide recommendation explanations with the aid of explicit topics. Secondly, we will propose the explainable recommendation models by integrating trust community detection and latent user/item factors learning, study the correlations between trust communities and latent factors, and provide recommendation explanations with the aid of trust communities. Thirdly, we will propose heterogeneous attributed graph-based explainable recommendation models, where the multi-source recommendation data is represented as a heterogeneous attributed graph with different types of vertices (users and items) and links (user-item preference, user-user social relation, item-item similarity), and each vertex having its own attributes. The main goal is to provide recommendation explanations by taking advantage of multi-source recommendation data. Fourthly, we will study the optimization and parallel computing theory, propose effective and efficient algorithm to solve the various explainable recommendation models. Finally, we will provide a multi-source recommendation system and test it on real-world applications. This project will show some new ideas and provide key techniques for explainable recommendation.

随着行业数据的急速增长，推荐系统成为用户获取有价值信息的必要工具。然而，已有系统着力于推荐准确率的提升，在很大程度上忽视了推荐系统的可解释性，进而影响用户的体验效果。本项目以设计精度高且可解释性强的推荐方法为目标，依据统计学习、协同学习、图计算、优化等理论，围绕推荐数据的多源、海量、动态等特点开展一系列研究。主要研究内容包括：研究物品信息主题与推荐隐变量的关联性，设计情感依赖的联合主题可解释性推荐模型，提供基于显性主题的推荐理由。研究用户信任团体与推荐隐变量的关联性，设计融合社交信息的协同矩阵分解可解释性推荐模型，提供基于信任团体的推荐理由。研究推荐数据的异构属性图表示方法，设计基于异构属性图的可解释性推荐模型，提供融合多源信息的推荐理由。研究优化理论和并行化思想，设计高效快速的模型优化算法及负载均衡的并行化算法。最终研制出融合多源信息的可解释性推荐系统，并结合实际应用平台加以验证。

随着行业数据的急速增长，推荐系统成为用户获取有价值信息的必要工具。然而，已有系统着力于推荐准确率的提升，在很大程度上忽视了推荐系统的可解释性，进而影响用户的体验效果。项目组重点分析推荐数据多源、海量、动态等特点，从理论研究和技术创新上取得以下成果：研究概率图模型和统计学习理论，探讨物品信息（如内容/评论信息等）对推荐的影响，重点研究信息主题、情感与推荐隐变量间的相互关联性，提出情感依赖的的联合主题可解释性推荐模型，提供基于显性主题的推荐理由；研究动态网络分析和矩阵计算理论，探讨社交网络对推荐的影响，重点研究动态用户信任团体构建思想、用户团体与推荐隐变量间的相互关系，提出融合社交信息的协同矩阵分解可解释性推荐模型，提供基于信任团体的推荐理由；研究信息表示和协同学习理论，分析多源推荐数据的协同性，设计基于异构属性图的多源推荐数据统一表示模型，提出基于异构属性图的可解释性推荐模型，确保模型构建时各信息源相互渗透，提供融合多源信息的相对健全的推荐理由；研究优化理论和并行化思想，探讨并行及增量计算中多源信息推荐数据划分、存储和计算结构，提出高效快速的可解释性推荐模型求解优化算法。在项目组成员的共同努力下，相继发表期刊论文27篇，会议论文27篇，接受论文6篇。包括SCI检索A1区期刊论文3篇，SCI检索A2区期刊论文6篇，SCI检索A3区期刊论文10篇，SCI检索A4区期刊论文3篇，；国际学术会议论文28篇（WWW, AAAI, ACL, ICCV, CVPR, ACM CIKM, ACM RecSys, KESM, PRCV, CCBR, ICDM, IJCNN, ISBDE, EMNLP, ICMI, BIG MM, ACML, ICCEA和 ACII）；国内EI检索学术期刊论文4篇；国内核心期刊论文6篇。培养15名博士研究生，41名硕士研究生，已毕业4名博士生，25名硕士生。