【作者】 洪明；

【导师】 王和勇；

【作者基本信息】 华南理工大学 ， 管理科学与工程， 2020， 博士

【摘要】 随着电子商务的蓬勃发展,网购平台聚集了众多的商家,为消费者提供多种多样的产品。面对日趋广阔的网络市场和日益激烈的同行竞争,研究网购平台如何优化其产品和服务,从而提高消费者的满意度和忠诚度,增加产品销量具有必要性和现实意义。在网购平台运营和发展的过程中,产生了海量的与消费相关的数据,包括产品评论、消费者投诉文本、广告业务数据。这些数据包含了网购平台中产品和服务的相关信息,因此,本论文从数据挖掘和文本挖掘的角度提出解决思路,为网购平台优化其产品和服务提供有效的技术、理论与方法。本论文的研究思路如下:(1)产品评论中包含的大量噪声词和无关词会带来很多噪声和无关的内容,而文本特征选择方法能够从中筛选出与产品优缺点相关的关键词,为生成产品评论摘要减少大量噪声。(2)面对海量的产品评论,商家需要花费大量的时间从中获得有关产品的重要信息,而产品评论摘要生成方法能够根据产品评论生成简洁易读的摘要,解决海量产品评论带来的“信息爆炸”问题。(3)网购流程的优化有利于提高消费者的购物体验,规范对商家的管理,减少购物纠纷,而目前有关网购平台服务质量的研究基本是从平台的整体入手,没有从网购流程的角度出发,研究其存在的问题并提出优化策略。基于消费者投诉文本,结合文本挖掘和服务科学模型的网购流程优化方法,既能在一定程度上避免了经验的限制,样本量不足和数据偏性等问题,又能反映消费者最关注的服务问题。(4)无效广告的投放为商家带来的收益远低于其投放成本,因此,提高有效广告识别率,帮助商家及时发现并移除无效广告,节省广告成本,是一个亟待解决的问题。基于客观的业务数据的广告分类模型,能够一定程度避免当前广告效果评价的研究中数据样本量不足,数据偏性,主观性强,量化困难等问题,有效解决有效广告和无效广告的分类问题。本论文的研究内容如下:(1)基于两种经典的深度学习模型卷积神经网络和长短期记忆网络,分别设计了对应的特征选择方法。首先,阐述提出的两种特征选择方法的具体原理,包括进行特征选择的步骤,深度学习模型的结构和训练。然后,基于多个来自网络的公开数据集,从分类性能、语义性能、存储性能三个角度,将提出的方法与传统的特征选择方法进行比较,验证方法的有效性。最后,基于分类性能、语义性能、存储性能三个角度的实验结果,对比分析提出的两种方法的性能差异。(2)结合词性规则、特征选择方法、主题模型和深度学习模型设计了产品评论摘要生成方法。第一,阐述产品评论摘要生成方法的具体原理,包括词性规则的设定,特征选择方法、主题模型和深度学习模型的作用。第二,利用从网络搜集的产品评论数据集对提出的方法展开实例分析,通过结果验证方法的有效性。(3)基于文本挖掘和服务科学模型设计了网购流程优化方法。第一,利用提出的产品评论摘要生成方法,从消费者投诉文本中挖掘网购流程存在的问题。第二,利用服务科学的模型“过程链网络”,结合发现的网购流程中的具体问题,提出网购流程优化的针对性策略。(4)提出了基于高斯滤波和决策树的广告分类模型。第一,阐述了分类模型的具体原理,包括高斯滤波对业务数据的作用,决策树的参数设置。第二,通过现实的广告业务数据验证了模型的有效性。本论文的创新点如下:(1)提出的基于深度学习的特征选择方法,为目前特征选择的研究提供了新的思路。将深度学习模型应用于特征选择,结合深度特征与传统的词频信息设计更有效的特征选择方法。对深度学习模型采取无指导的训练方法,提高了方法对无指导文本的适用性。(2)提出的综合多种方法优势的产品评论摘要生成方法,丰富了目前产品评论内容挖掘的研究。结合词性规则、主题模型和特征选择设计了三层信息提取机制,保证提取的信息能够反映产品评论的重要内容。利用特征选择方法提供了一种交互机制,用户能够从特征选择提供的关键词中选择其需要了解的产品属性。依托深度学习模型“长短期记忆网络”自主学习文本信息以及文本预测的优势,从提取的信息中学习关键信息,自动生成摘要。(3)提出的结合文本挖掘和服务科学模型的网购流程优化方法,扩展了当前网购平台服务质量的研究。将提出的产品评论摘要生成方法和服务科学模型运用到网购流程优化中,对消费者最关注的网购流程问题展开讨论和优化,避免人为经验的限制和调研方法中可能存在的样本不足或数据偏性等问题。(4)提出的基于业务视角的广告分类模型,为当前广告效果评价的研究提供了新的思路。采用高斯滤波调整原始业务数据的分布,缓解特征工程、主成分分析等方法处理业务数据都无法避免的有效和无效广告在分类过程中由于特征不明显而难以区分的问题。利用C5.0决策树构建了分类器,提高有效广告识别率,帮助商家尽早移除无效广告,降低成本。综上,网购平台(商家)可以从产品评论、网购流程和广告三个方面优化其产品和服务,提高消费者的满意度和忠诚度,增加产品销量,具体的建议包括:定期分析产品的优缺点以及时调整销售策略,定期分析和优化网购流程以提供高效的服务,以及定期优化广告投放策略以提高消费者的购物体验。更多还原

【Abstract】 With the rapid development of E-commerce,online shopping platforms gather lots of sellers and provide customers various kinds of products.Under the background of larger online market and more severe competition,it is of necessity and practical meaning to study how online shopping platforms improve products and services,in order to improve the satisfaction and lyalty of customers.In the operation and development of onlne shopping platforms,large quantities of cumsuming-related data are created,including product reviews,customer complaint text and advertisement business data.These various kinds of data contain related information of consuming.This essay proposes solutions in the aspect of data mining and text mining to provide effective techniques,principles and methods.The research route of this essay is shown below.(1)Product reviews conain large quantites of noisy and irrelevant words,resulting in a lot of noisy and irrelevant contents.Feature selection methods are capable of selecting keywords which are related to advantages and disadvantages of products to reduce much noise for generating produt review adstracts.(2)Facing large quanties of product reviews,sellers need to spend a lot of time to obtain importat information of products.The method for generating product review abstracts is able to generate clear and readable abstracts to solve the problem of information explosion.(3)The optimization of online shopping processes is helpful to raise purchasing experiences of customers,regulate seller management and reduce purchasing conflicts.However,current researuches which are related to service quality of online shopping platforms generally focus on the whole platforms,ignoring the existing problems and optimizatio n strategies of online shopping processes.Based on customer complaint text,the optimization method for online shopping processes combining text mining and service science model is capable of avoiding the problems of experience limitation,lack of samples and data biases on the one hand,and reflecting the most customer-concerned service problems.(4)The display of ineffective advertisements brings far less profits compared to the ir costs,so it is urgent to find the solution for raising the recognition rates of effective advertisements and helping sellers discover and remove ineffective advertisements in time to save costs.Advertisement classification model based on objective business data,to some extence,is able to avoid the problems such as lack of data samples,data biases,strong subjectivity,quantification difficulties in current researches on advertisement effectivenss evaluation,in order to solve the problem of effective and ineffective advertisements classification.The research contents of this essay are shown below.(1)Two feature selection methods are designed based on two typical deep learning models,that is,convolutional neural network and long-short term memory network.To begin with,the specific principles of the two methods are described,including steps of performing feature selection,the structures and training of the deep learning models.Next,the two proposed methods are compared with traditional methods in the three aspects of classification performance,semantical performance and storage performance on their effectivenss,based on several public datasets.Finally,performance of the two proposed method s is compared based on the experimental results of classification performance,semantical performance and storage performance.(2)Method for generating product review abstracts is designed combining part-of-speech(POS)rules,feature selection method,topic model and deep learning model.Firstly,the specific principles of the method is described,including the setting of POS rules,the roles of feature selection method,topic model and deep learning model.Secondly,datasets of product reviews from web are used to perform case studies to validate effectiveness of the method.(3)Optimization method for online shopping processes is designed based on text mining and service science model.Firstly,the proposed method for generating product review abstracts is used to mine the existing problems of online shopping processes from customer complaint text.Secondly,the service science model ―process chain network ‖ is used to propose targeted optimization strategies,according to the mined specific problems of online shopping processes.(4)Advertisement classification model based on Gaussian filter and decision tree is proposed.Firstly,the specific principles of the model are descr ibed,including the affect of Gaussian filter on business data and the parameter settings of decisioin tree.Secondly,the effectiveness of the model is validated by using practical advertisement business data.The innovations of this essay are shown below.(1)The proposed methods based on deep learning provide new research route for current researches on feature selection.The deep learning models are used in feature selection to designed more effective methods by combining deep features and traditional term frequency information.Unsupervised training strategy is selected for the training of deep learning models to improve the applicability for unsupervised text.(2)The proposed method for generating produc t review abstracts combining advantages of several methods enriches the researches on current researches on product reviews mining.A three-level information extraction mechanism is designed to guarantee that the extracted information is able to reflect important contents of product reviews.An interactive mechanism is provided by feature selection for customers to select更多还原