【作者】 王超；

【导师】 熊辉； 陈恩红；

【作者基本信息】 中国科学技术大学 ， 计算机应用技术， 2022， 博士

【摘要】 个性化推荐技术是指在线上推荐应用场景中向使用系统的用户推荐符合自身兴趣偏好和意图的产品。近年来,随着互联网相关技术和产业的快速发展,推荐系统被广泛应用于线上购物平台、新闻媒体推送、视频音乐网站等在线服务场景中。推荐系统一方面可以改善用户体验、提高用户粘性,另一方面也可以增加产品曝光量、提升内容分发效率,具有极高的商业价值。因此,推荐系统成为了计算机、管理学及相关交叉学科中的一个长期热门研究问题,受到了众多研究者的关注。然而,现有的推荐系统仍然面临着用户决策意图和特征画像构建困难、单项决策和序列决策中的偏好建模复杂多变、推荐结果可信程度无法量化等挑战。管理学中将用户的决策过程分为五个阶段:需求确认、信息收集、备选产品评估、购买决策和决策评估。推荐系统为提升用户在这五个决策阶段中的用户体验,需要面临三个主要的研究问题,分别是用户理解、决策理解和决策评估。为此,本文利用机器学习、数据挖掘等技术,围绕用户决策过程中的五个阶段,在三个主要研究问题上系统性地开展了面向个性化推荐的相关技术和应用研究工作,提升用户体验。针对这五个阶段,分别提出了基于自编码器的用户意图画像建模方法、基于主题模型的用户特征画像建模方法、基于集合偏序对比的决策偏好建模方法、基于时间感知的序列决策偏好建模方法和基于贝叶斯估计的决策置信度建模方法。本文的主要工作与贡献可以概括如下:首先,构建全面的用户理解是分析用户决策行为和推荐的基础,为此本文提出建模用户的意图画像和特征画像等方法。一方面,在需求确认阶段,本文提出了一种基于自编码器的用户意图画像建模方法。传统推荐系统建模中用户意图和用户偏好这两种因素被耦合在一起建模,为解决此问题,本文采取双重解耦的设计。第一重解耦用于在建模过程中分离用户的意图因素和偏好因素,第二重解耦则是对一个用户的不同决策意图分别构建独立的用户表征向量,实现细粒度的用户意图和偏好刻画。最终在三个真实数据集上进行了大量实验,验证了提出的双重解耦表征学习方法在意图识别和推荐任务上的有效性和可解释性。另一方面,在信息收集阶段,本文提出了一种基于主题模型的用户特征画像建模方法。标签数据是推荐系统中最常见的一种特征类型,然而标签数据容易存在稀疏性、错漏性和变化性等问题,如何利用标签数据来构建用户特征画像是一个关键又困难的任务。本文利用一个非常有代表性的基于标签数据进行推荐的应用场景,即个性化员工在线课程推荐场景,进行了相关研究。利用技能标签数据联合建模员工的当前能力和职业发展偏好,并确保学到的用户能力画像和职业发展画像都具有很高的可解释性。最终基于真实的企业内部课程平台应用场景及收集与平台上的真实数据进行了实验验证,验证了用户特征画像的有效性、可解释性,以及推荐算法的准确性、鲁棒性。其次,构建深度的决策理解是分析用户决策行为和推荐的核心,为此本文提出对单次决策和多次决策的精准建模方法。一方面,在备选产品评估阶段,本文提出一种基于集合偏序对比的决策偏好建模方法。传统的用户决策偏好建模可以分为逐点建模方法、成对对比建模方法和列表排序建模方法。然而在实践中,成对对比建模方法的独立性假设往往难以满足。此外,逐点建模方法和列表排序建模方法都不能有效地处理未观测样本数据。本文提出的集合偏序对比建模方法可以更好地拟合推荐系统中的数据特点,为决策偏好建模问题开拓了新思路。最终在三个真实世界数据集上的大量实验清楚地验证了集合偏序对比建模方法的准确性和有效性。另一方面,在购买决策阶段,本文提出一种基于时间感知的序列决策偏好建模方法。在时间序列等序列决策场景中,用户的决策行为之间存在连续性和依赖性,必须考虑到多决策之间的关联影响。本文基于一个典型的时间序列多任务决策问题,即员工跳槽问题展开了研究。为此,本文采用深度时间序列建模与多任务协同过滤相结合的方式,借助所提出的时间感知模块,可以在预测阶段输入不同的跳槽时间来生成不同的预测结果,实现推荐结果随时间变化而变化。最终基于从真实职业社交网站上收集的简历数据集,进行了大量实验验证了方法的有效性。最后,在决策评估阶段,需要量化评估决策的好坏,例如可信程度等指标。为此,本文提出一种基于贝叶斯估计的决策置信度建模方法。传统推荐只提供预测的评分值,而不能量化这次预测的可信程度,本文提出的基于贝叶斯估计的用户决策置信度建模框架可以在保障评分预测的准确性的同时测量预测的置信度。具体而言,在矩阵分解模型中引入不同用户和产品的方差参数,通过预测置信区间来度量评分预测的置信度。由此,用户和产品对评分的置信差异影响都被考虑在建模内,这种联合评分预测和置信度预测的建模过程同时也能提高评分预测任务的准确性。最终在真实数据集上进行了大量实验验证,在评分预测和置信度量化两个任务上验证了方法的有效性。更多还原

【Abstract】 Personalized recommendation technology refers to recommending products that meet users’ own interests,preferences and intentions in online recommendation application scenarios.In recent years,with the rapid development of Internet-related technologies and industries,recommender systems have been widely used in online service scenarios such as online shopping platforms,stream media apps,and music websites.On the one hand,the recommender systems can improve user experience and stickiness.On the other hand,it can also increase product exposure and improve the efficiency of content distribution,which has extremely high commercial value.Therefore,recommender system has become a long-term popular research problem in computer science,management and related interdisciplinary subjects.It has attracted the attention of many researchers.However,the existing recommender systems still face the difficulty of understanding the user’s intentions and portraits,the complex and changeable preference modeling,and the unquantifiable recommendation confidence evaluation.In management filed,the decision-making process of a user is divided into five stages:need recognition,information search,alternatives evaluation,purchase decision and decision evaluation.To improve the user experience in these five decision-making stages,the recommender system needs to face three main research problems:user understanding,decision understanding,and decision evaluation.To this end,this paper uses machine learning,data mining and other technologies to systematically perform research for personalized recommendation around the five stages in the user decision-making process to improve user experience.For these five stages,the disentangled user decision-making intention modeling method,user portrait modeling method,setwise preference ranking modeling method,time-aware sequential decision modeling method,and user decisionmaking confidence modeling method are respectively proposed.The main work and contributions of this paper can be summarized as follows:First,it is important to construct comprehensive user understanding through modeling user intentions and portraits.On the one hand,in the need recognition stage,we propose a user intention modeling approach.In traditional collaborative filtering approaches,both intention and preference factors are usually entangled in the modeling process,which significantly limits the robustness and interpretability of recommendation performances.To this end,we propose a double disentangled collaborative filtering(DDCF)approach for personalized recommendations.The first-level disentanglement is for separating the influence factors of intention and preference,while the second-level disentanglement is performed to build independent sparse preference representations under individual intention with limited computational complexity.Finally,extensive experiments on three real-world datasets clearly validate the effectiveness and the interpretability of DDCF.On the other hand,in the information search stage,we present a focused study on the explainable user portrait modeling method.We take the employee training course recommendation problem as an example to show how to perform user portrait modeling.Specifically,we jointly model both the employees’ current competencies and their career development preferences in an explainable way.First,we extract the latent interpretable representations of the employees’ competencies from their skill profiles with autoencoding variational inference based topic modeling.Then,we develop an effective demand recognition mechanism for learning the personal demands of career development for employees.Finally,we can generate explainable recommendation results based on the competency representations.Extensive experimental results on real-world data clearly demonstrate the effectiveness and the interpretability of both of our frameworks,as well as their robustness on sparse and cold-start scenarios.Second,it is important to construct deep decision understanding for single and multiple decisions.On the one hand,in the alternatives evaluation stage,we study the preference ranking modeling problem.While considerable efforts have been made in this direction,the well-known pairwise and listwise approaches have still been limited by various challenges.Specifically,for the pairwise approaches,the assumption of independent pairwise preference is not always held in practice.Also,the listwise approaches cannot efficiently accommodate items with the same rating value and unobserved data due to the precondition of the entire list permutation.To this end,we propose a novel setwise Bayesian approach for collaborative ranking,namely SetRank,to inherently accommodate the characteristics of user feedback in recommender systems.Finally,extensive experiments on three real-world datasets clearly validate the superiority of SetRank compared with various state-of-the-art baselines.On the other hand,in the purchase decision stage,we propose a time-aware sequential decision-making modeling approach.The dependence between multiple decision sequences must be considered in the decision-making process.Here we take the career trajectory prediction problem as an example.Specifically,we first exploit a hierarchical deep sequential modeling network for career embedding.Furthermore,we propose a temporal encoding mechanism to handle dynamic temporal information so that we can generate time-aware predictions by addressing the challenges for variable interval time sequence modeling.Finally,we have conducted extensive experiments on large-scale real-world data,and the results show that our approach has advantages on all tasks.Third,in the decision evaluation stage,we propose a Confidence-aware Matrix Factorization(CMF)framework to simultaneously optimize the accuracy of rating prediction and measure the prediction confidence in the model.Specifically,we introduce variance parameters for both users and items in the matrix factorization process.Then,prediction interval can be computed to measure confidence for each predicted rating.These confidence quantities can be used to enhance the quality of recommendation results based on Confidence-aware Ranking(CR).We also develop two effective implementations of our framework to compute the confidence-aware matrix factorization for large-scale data.Finally,extensive experiments on three real-world datasets demonstrate the effectiveness of our framework from multiple perspectives.更多还原