融合多维度异质资源的推荐解释自动生成模型研究

Recommendation explanation is one of the three core problems in recommendation systems, and has been a challenging problem to support many applications in academia and industry for a long period of time. An intuitive and convincing recommendation explanation could not only improve the user acceptance to the recommendation results, but also enhance the user experience of recommendation systems in many aspects. Therefore, on the basis of heterogeneous and massive data, this study aims to provide explainable recommendation results in the form of probability and semantic, and automatically generate multi-dimensional explanations: Firstly, we explore some existing methods, such as sparse learning, bayesian network, artificial immune, deep learning, to mine user, product and context which provide knowledge sources to generate recommendation explanations. Secondly, in order to reduce the complexity of the explanation model, we intend to identify the users' recommendation style using deep reinforcement learning, and regard recommendation style as the basic structure of multi-dimensional recommendation explanation model. Finally, we utilize probabilistic graphical model and graph mining method for heterogeneous networks to construct the automatic generation model of recommendation explanation. The final goal is to generate explanations in the form of probability and semantic by fusing multi-dimensional heterogeneous features. The study meets the personalized, dynamic and multi-dimensional modeling for recommendation explanation. It overcomes the shortcomings of traditional recommendation explanation model that utilizes single mode of explanation and ignores the correlation between recommendation results and recommendation explanations. It gives novel ideas and theoretical principles for further research on the automatic generation of recommendation explanation, and is a research with important academic significance and great application value.

推荐解释是推荐系统需要解决的三大核心问题之一，是长期以来困扰学术界和企业界的重要问题。直观、可信服的推荐解释不仅能够提高用户对推荐结果的接受度，还有助于在诸多方面增强用户体验。因此，本项目立足异质海量数据，以概率和语义形式解释推荐结果并自动生成多维度的推荐解释：首先，拟利用稀疏学习、贝叶斯网络、人工免疫和深度学习等方法挖掘用户、商品和情境数据,为生成推荐解释提供知识源；然后，拟通过深度强化学习方法识别用户的推荐解释风格，将解释风格作为多维度推荐解释的基本结构以降低解释模型的学习复杂度；最后，拟借助概率图模型和异质网络图挖掘方法，融合多维度异质特征以概率和语义形式解释推荐结果，最终实现推荐解释自动生成。本项目考虑了推荐解释的个性化、动态性和多维度建模，克服了传统的推荐解释模式单一以及推荐解释与推荐算法毫无联系的缺点，为推荐解释自动生成研究提供了新思路和理论依据，具有重要的学术意义和应用价值。

推荐解释是推荐系统需要解决的三大核心问题之一，是长期以来困扰学术界和企业界的重要问题。直观、可信服的推荐解释不仅能够提高用户对推荐结果的接受度，还有助于在诸多方面增强用户体验。因此，本项目立足异质海量数据，以用户和商品的属性关系以及二者之间交互关系作为情境解释推荐结果并自动生成多钟解释风格的推荐解释。主要研究内容包括：a) 围绕推荐系统中用户情境信息开展研究，包括面向推荐的用户兴趣扩展方法、推荐系统中信任网络构建、基于相似度投票的社区划分改进算法、面向重要节点的社区发现算法和基于社区结构信息的网络表示学习等。b) 围绕推荐系统中商品评论开展研究，包括采用词向量注意力机制的双路卷积神经网络句子分类模型、基于稀疏自学习卷积神经网络的句子分类模型和面向双注意力网络的特定方面情感分析模型。c) 围绕推荐系统和推荐解释开展研究，包括查分隐私和时序的推荐模型、基于边信息的用户和商品自适应深度建模的推荐、基于层次化相似性度量的深度隐性因子模型的推荐系统、基于社会关系的贝叶斯个性化排序隐性反馈的推荐模型、融合评论文本细粒度情感的快速可解释推荐、面向细粒度数据的可解释推荐链模型以及面向注意力向导游走的多风格推荐解释自动生成模型等。在Epinions、Ciao、Amazon、Yelp、Movielens、Citation Network等多个推荐系统数据集上，项目组提出的模型和方法在多项评测标准中明显优于其它的基本方法。本项目考虑了推荐解释的个性化、动态性和多维度建模，克服了传统的推荐解释模式单一以及推荐解释与推荐算法毫无联系的缺点，为推荐解释自动生成研究提供了新思路和理论依据，具有重要的学术意义和应用价值。