基于多标记学习的用户属性流式预测模型研究与实现

发布时间：2018-10-04 23:13

【摘要】：互联网正在从“用户以获取信息为主”的Web1.0时代过渡到“用户既是网络信息获取者又是网络信息制造者”的Web2.0时代。为了能够在海量数据中发掘信息或者服务,用户画像具有巨大的作用和价值,能够为个性化搜索、个性化推荐、广告营销、产品策略以及运营方向提供基础性支持,指引方向。用户属性预测是用户画像研究的核心工作,如今用户属性预测的研究主要着重于单个属性的预测模型构建,缺乏较为完善的、全面的多个属性同时预测的模型方法;此外,更是缺乏相应领域的数据流挖掘以及概念漂移处理机制,无法实现用户属性动态预测,并且现有的概念漂移研究具有局限性,需要进行相应地改进与加强。本文从以上问题入手,旨在构建体系完备、效率与性能优越的用户属性流式预测模型。在属性预测方面,本文侧重同时预测多个属性的理念,基于多标记学习技术,采用多示例多标记框架(MIML)将属性预测作为一个广义的多标记分类研究,并且创新性构建组成用户对象的示例概念,采用聚类方法构建示例,实现了能够快速、准确、同时预测多个属性的模型构建。不同于离线预测模型,本文创新性加入基于数据流挖掘技术的在线流式框架,处理用户产生的在线行为与动态,侧重处理数据流的各类概念漂移问题,提出了一种基于原型(Prototype-based)学习的自适应概念漂移分类算法SyncPrototype,相较于现有算法,SyncPrototype在分类性能、概念漂移的响应速度以及时间性能等方面都有明显提高,能够更加有效处理并适应数据流概念漂移问题。为用户属性流式增量迭代方面提供了有力支持,从而实现用户属性动态预测以及流式迭代。本文运用基于多标记学习的用户属性流式预测模型设计开发了用户属性认证系统的数据挖掘验证模块,能够有效验证微博用户所填个人信息真实性,衡量属性可信度。
[Abstract]:The Internet is transitioning from the era of Web1.0, where users are mainly to get information, to the era of Web2.0, where users are both network information acquirers and network information makers. In order to discover information or service in mass data, user portrait has great function and value. It can provide basic support and direction for personalized search, personalized recommendation, advertising marketing, product strategy and operation direction. User attribute prediction is the core work of user portrait research. Nowadays, the research of user attribute prediction mainly focuses on the construction of a single attribute prediction model, which lacks a more perfect and comprehensive model method for simultaneous prediction of multiple attributes. There is also a lack of data stream mining and conceptual drift processing mechanism in the corresponding fields, which can not realize the dynamic prediction of user attributes, and the existing research on concept drift has limitations, so it needs to be improved and strengthened accordingly. The purpose of this paper is to construct a user attribute flow prediction model with complete system, high efficiency and superior performance. In the aspect of attribute prediction, this paper focuses on the concept of simultaneous prediction of multiple attributes. Based on multi-label learning technology, this paper uses multi-example and multi-label framework (MIML) to study attribute prediction as a generalized multi-label classification. The concept of user object is innovatively constructed and the example is constructed by clustering method. The model can be constructed quickly and accurately and can predict multiple attributes at the same time. Different from the offline prediction model, this paper creatively adds an online flow framework based on data stream mining technology to deal with the online behavior and dynamics generated by users, and focuses on dealing with various conceptual drift problems of data flow. An adaptive concept drift classification algorithm based on prototype learning (Prototype-based) is proposed. Compared with the existing algorithm, SyncPrototype, has a significant improvement in classification performance, response speed and time performance of concept drift. It can deal with and adapt to the concept drift problem of data flow more effectively. It provides powerful support for user attribute flow incremental iteration, so as to realize user attribute dynamic prediction and flow iteration. In this paper, the user attribute flow prediction model based on multi-label learning is used to design and develop the data mining verification module of user attribute authentication system, which can effectively verify the authenticity of personal information filled by Weibo user and measure the reliability of attributes.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP393.0;TP311.13

【相似文献】