CASIA OpenIR  > 毕业生  > 博士学位论文
基于深度表达学习的用户建模研究
刘强1,2
学位类型工学博士
导师王亮
2018-05-24
学位授予单位中国科学院大学
学位授予地点北京
关键词用户建模 深度表达学习 个性化推荐 上下文感知 序列行为预测 不实信息检测
摘要

近年来,随着互联网技术的飞速发展,社交媒体、在线购物网站等越来越受到人们的欢迎,人们可以从这些丰富的网络应用中寻找各自感兴趣的内容或需要的物品。以用户为中心的网络数据包罗万象,除了数量巨大,形式也是多种多样,有文本、图像、视频、关系、行为等等,如何从如此纷繁复杂的信息中挖掘出有价值的信息是个重要的挑战。利用这些信息,对用户的兴趣、行为进行理解,对用户生成的信息进行分析,进行用户建模的研究,构建用户全方位的表达,有着重要而实际的意义,可以在个性化推荐、在线广告、异常信息检测等方面产生推动性的成果。目前,深度表达学习的研究在图像、语音等多个领域带来了巨大的提升,以此方法进行用户建模,构建用户表达,是一个十分有效而具前景的方法。本论文旨在基于深度表达学习方法,对用户建模中的几个关键性问题进行针对性的研究,提出对应的解决和改进方案,具体研究如下:

 

(1) 用户在网络上的行为有着丰富的上下文信息,包括时间、地点、天气等情境信息。在不同的情境下,用户有着不同的行为模式。学习用户在不同上下文下的表达,有着很强的应用价值,可以有效实现用户行为的预测,应用在个性化推荐、点击预测等场景。传统的基于因子分解的上下文建模方法,关注的是实体间的相关性和相似度,难以挖掘实体间的深层联系,无法建模实体间的状态变换作用。因此,我们提出了上下文信息对用户和目标物品所具有的操作特性,并用张量乘法的形式进行建模。随后,我们将模型进一步一般化,扩展到更复杂的多实体交互场景,并用递归结构进行建模。我们的方法在上下文感知推荐、点击预测等应用中取得了优于之前方法的效果。

 

(2) 用户行为具有很强的时序依赖性,即之前的行为信息对之后的行为有着很大的影响,这种特性对于个性化建模有着重要作用。传统的基于马尔可夫链的行为预测方法,由于其强独立性和无法建模行为间隐含关系的缺点,逐渐被循环神经网络取代。相比于循环神经网络传统应用的文本、视频等数据,用户行为数据包含更复杂的信息而难以被模型直接建模,如外部情境、时间差、局部行为模式、多行为信息等。因此,针对这些信息,我们对循环神经网络的结构进行了适配与优化。我们的方法在多个用户行为预测的数据集上都取得了当下最优的效果。

 

(3) 伴随着用户的行为,还有大量的多模态内容信息,如商品的图像信息。在用户购买衣服、装饰等商品时,这些图像信息往往起到了很关键的作用,为基于视觉的个性化推荐提出了需求。针对于这个问题,我们提出了一种风格感知的神经网络,可以建模商品的风格属性,把握用户的兴趣偏好。我们在真实数据上对我们的方法进行了实验和可视化聚类,可有效地把握商品和用户的风格特性,取得了优于之前方法的效果。

 

(4) 在网络应用中,除了上述用户正常的使用行为,还往往存在一些异常的行为,如不实信息的发布与传播。这些不实信息的传播对互联网和社会的健康发展有着很大的危害,如何实现事件真假的自动快速的检测,是一个十分重要的问题。网上与一个事件相关的信息往往有很多条,而其中只有很小的一部分对检测能起到很重要的作用,这些有用的信息很容易被大量的无用信息淹没掉。因此,我们提出了一种基于注意力机制的不实信息检测方法,根据文本内容和动态时间两方面信息,自动抽取对检测起到重要作用的信息,提升检测效果。我们的方法在多个社交媒体数据集上都取得了当下最优的效果。

其他摘要

Recently, with the growth of the Internet, online platforms such as social media and online shopping website are getting more and more popular. People can access to what they like and need on these websites. Online data, which is user-centered, has great amount, as well as varied forms, e.g., text, image, relation and behavior. It is a challenge to mine valuable information from these complex data. Understanding user interest and behavior, analyzing user data, and conducting research on user modeling has great practical meaning. Such research can promote the applications of personalized recommendation, online advertising, abnormal detection and so on. Currently, deep representation learning has succeeded in areas such as text, image and video processing. Based on deep representation learning, it is promising to conduct research on user modeling and constructing user presentations. In this paper, based on deep representation learning, we plan to provide solutions for several key problems in user modeling. Our contributions are listed below:

 

(1) User behaviors are usually associated with a variety of contextual information, e.g., time, date, location and weather. In different contexts, user usually has different behavioral patterns. It is practical to learn context-aware representations of users, which can be applied in behavior prediction, personalized recommendation and click prediction. Conventional factorization-based methods focus on modeling the correlations and similarities among entities. This fails to capture the deep connections among entities, and model the operating effects of some entities. Accordingly, we address the operating effects of contextual information on users and items, which can be modeled by contextual operating tensors. Furthermore, we generalize the situation to multi-entity interaction, which is modeled via a recursive structure. Our methods achieve state-of-the-art performances in context-aware recommendation and click prediction.

 

(2) There is strong temporal dependency among behaviors of a user. This shows great importance on predicting next behaviors and personalized recommendation. The conventional Markov chain-based methods for sequential behavior prediction are recently replaced by the advanced recurrent neural networks. However, there are some kinds of information in user modeling scenarios that can not be modeled by conventional recurrent neural networks: external contexts, time intervals, local behavioral patterns and multiple behaviors. Accordingly, to capture these information, we optimize the structure of recurrent neural networks. Our methods achieve state-of-the-art performances on several user behavior prediction datasets.

 

(3) User behaviors are also associated with some multimodal content information, such as the visual information of items. When users are purchasing products like clothes and decorations, visual information plays a key role. This makes visual recommendation practical and necessary. To solve this problem, we propose novel style-aware neural networks called DeepStyle. It can model the style features of items, and capture the preferences of users. We conducted experiments and visualized the clustering results of style features on real datasets. The results show that our method achieves the state-of-the-art performances in visual recommendation.

 

(4) Besides above normal user behaviors, there also exists some abnormal behaviors on the Internet. Among them, the posting and spreading of misinformation has great harm to the Internet and the society. It is a crucial task to automatically detect whether an event on the Internet is true or not. On the Internet, an event usually consists of a great number of messages, only a small part of which are important to the detection of misinformation. Thus, useful information can be easily overwhelmed by useless information. Accordingly, we propose a novel misinformation detection method based on the attention mechanism. It can automatically select messages that are most important to misinformation identification. Experiments conducted on several real-world social media datasets show that our method promotes the performance of misinformation identification.

文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/21013
专题毕业生_博士学位论文
作者单位1.中国科学院自动化研究所
2.中国科学院大学
推荐引用方式
GB/T 7714
刘强. 基于深度表达学习的用户建模研究[D]. 北京. 中国科学院大学,2018.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
签字版本.pdf(6011KB)学位论文 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[刘强]的文章
百度学术
百度学术中相似的文章
[刘强]的文章
必应学术
必应学术中相似的文章
[刘强]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。