【作者】 孙丽；

【导师】 曹锦丹；

【作者基本信息】 吉林大学 ， 社会医学与卫生事业管理， 2015， 博士

【摘要】 【研究意义与目的】根据世界卫生组织（WHO）研究报告：人类三分之一的疾病通过预防保健可以避免；三分之一的疾病通过早期发现可以得到有效控制；三分之一的疾病通过信息的沟通可以提高治疗效果。可见，健康信息资源对卫生保健、疾病提早干预及辅助治疗有着重要作用。由于互联网信息的海量、多元化和查询便捷等特点，已经成为获取健康信息资源的重要渠道，虽说其不是获取医疗健康信息资源的正规机构，但可以是一个医疗保健知识的补充，无疑可以给患者及用户一个深入了解卫生保健、疾病预防和辅助治疗的信息窗口，增强对健康的意识及管理。有学者研究，越常网上搜寻健康信息，越利于健康行为的改变，利于健康决策能力的提高。近年来，越来越多的用户网上搜寻健康信息，健康信息需求不断增长，而健康素养水平又普遍偏低，且目前我国健康网站服务方式缺乏个性化，因此，急需进一步拓展和提升网络健康信息服务方式和质量，提供有针对性的健康信息服务，提高用户健康信息搜寻的效率及满意度。研究用户的网络健康信息搜寻行为，探明规律，进而提供有针对性的搜寻帮助，对完善网络个性化服务方式是非常必要且有意义的。任务是用户进行信息搜寻的原动力，是影响信息搜寻行为的主要因素。对任务的研究会有助于理解人们为什么搜寻信息、搜寻信息的类型及选择、获取和使用信息的方法。因此，查明任务和信息搜寻行为的关系是解决问题的思路。目前，研究普遍关注任务对信息搜寻行为的影响，而如何据此进行更有针对性的个性化服务，这个领域还有待拓展，本研究在此进行深入推进，通过对用户完成不同任务时，网络信息搜寻过程展开系统而深入的研究，查明任务和信息搜寻行为的影响关系，进而通过信息搜寻行为动态预测任务类型，再根据不同任务类型下，用户信息搜寻行为的规律性进行个性化服务策略开发。【研究方法】本研究采取实验研究方法，数据通过调查问卷及屏幕录像软件获得。实验共邀请101名大学生、研究生作为调查对象，获得有效问卷101份，完成有效健康信息搜寻任务606个。数据分析采用统计学数据分析方法。预测模型通过数据挖掘方法构建。【研究内容和结果】（1）任务特征与网络健康信息搜寻行为影响关系研究①任务框架构建本文提出的任务框架构建围绕任务类型展开，而任务类型划分受制于任务特征的选取，因此重点对任务特征进行系统分析，对任务特征的界定和度量方法进行深入阐述，在此基础上再确定任务类型，探讨其与信息搜寻行为之间的影响关系，这样可以形成一个明确的研究框架，利于问题的深入分析。任务特征是指任务本身所具有的或者外部环境所限定的特征，主要体现在任务的内部特征和外部特征两个方面。任务的内部特征指的是任务独立于外部环境以及任务执行者的特点，任务内部特征分为基本任务特征和复合任务特征，在基本任务特征基础上，再抽象出复合任务特征，这样便于特征的界定和度量。基本任务特征可以设定为任务的目标特征、路径特征和结果特征3大部分。复杂性和结构性是影响信息搜寻行为重要的复合任务特征。本研究以任务基本特征中的目标特征为基础划分任务类型，将健康领域内任务划分为事实型任务、信息扩展型任务和决策型任务三种。任务的外部特征是指独立于任务本身，由外部环境所决定的任务特点。在任务的环境属性中，就信息搜寻层面来讲，任务执行者特征对信息搜寻行为具有重要影响。任务执行者特征中已有知识与认知特征被认为是影响信息搜寻行为的重要因素，此外调查了人口学特征对健康信息搜寻行为的影响。已有知识包括健康领域知识和网络知识，采用健康信息素养量表获取。个体的认知差异主要体现于认知需求和认知风格维度上，分别通过认知需求量表和认知风格镶嵌图形测验量表获取。人口学特征包括性别、文化程度，通过调查问卷获得。任务外部特征中，任务情境（网络信息系统）对网络健康信息搜寻行为的影响，主要考查网站结构和网页内容两个方面。②信息搜寻行为变量信息搜寻行为变量的选取围绕信息搜寻流程来设计，包括：构建查询式方面变量（构建首个查询式的时间、查询长度、查询方式）、结果列表评估方面变量（唯一搜索引擎结果页面（SERPs）数量、所有SERPs总数量、唯一SERPs第一次停留平均时间、所有SERPs停留总时间、SERPs平均停留时间）、目标网站/网页操作方面变量（保存页面总数量、浏览页面总数量、浏览唯一页面数量、唯一页面第一次停留平均时间、页面停留总时间、页面平均停留时间、唯一页面第一次停留平均时间/唯一SERPs第一次平均停留时间、页面停留总时间/所有SERPs停留总时间）、结果不满意时操作方面变量（修改查询式（查询重构）次数、更换查询系统或更改查询的限定范围次数、从结果列表中再次查询次数）及任务总体方面变量（任务完成时间、网站会话数）。此外，由于本研究的任务（如信息扩展型任务、决策型任务）结果很难用正确与否来衡量，故而，绩效测量采用主观评价的方式来辅助衡量，变量包括：付出努力程度、满意度。③任务特征与信息搜寻行为影响关系经数据统计分析显示：任务内部特征（目标特征）对网络健康信息搜寻行为具有重要影响，信息搜寻行为变量中，除查询方式、更换查询系统次数不受任务类型影响外，其余变量均具有显著差异性。任务外部特征中，任务执行者特征的认知需求对所有SERPs停留总时间、页面停留总时间有重要影响；健康信息素养对所有SERPs总数量、唯一页面第一次停留平均时间有重要影响；文化程度对页面停留平均时间有重要影响；而性别和认知风格对信息搜寻行为变量不具重要影响。任务外部特征中，网络信息系统的网站导航清楚性、网站链接和运行速度、网站可获取性、网站目录分类合理性对信息搜寻行为影响较大；网页内容的准确性、权威性、易获取性、客观性对信息搜寻行为影响较大。（2）网络健康信息搜寻任务类型预测研究任务类型会影响用户如何和为何选择、发现和评估信息资源，因此，如果系统能够监测用户正在进行的信息搜寻任务类型，就可以为用户信息搜寻提供有针对性的帮助，进而优化系统。此部分通过对用户网络信息搜寻行为与任务关系的分析，探明规律，通过数据挖掘方法，建立任务类型预测模型。本研究中数据挖掘的主要目标是根据网络信息搜寻行为变量，构建任务类型预测模型，属于分类预测。①基于整个任务期模型预测输入特征向量选择：根据任务特征变量对网络信息搜寻行为的影响结果，筛选适合的输入特征变量。不同任务类型下，信息搜寻行为变量存在显著差异性，则进入模型，此外，信息搜寻行为变量如受其他因素影响，则说明存在噪声，会考虑予以剔除。经上述分析，筛选后的信息搜寻任务类型预测指标有：完成时间、网站会话数、首次构建查询式时间、查询式关键词个数、唯一SERPs第一次停留平均时间、唯一SERPs数量、SERPs停留平均时间、保存页面总数量、浏览页面总数量、浏览唯一页面数量、唯一页面第一次停留平均时间/唯一SERPs第一次停留平均时间、页面停留总时间/所有SERPs停留总时间、结果列表再次查看次数、查询重构次数共14个变量。分类器的构造方法及评价：本研究采用SPSS Clementine进行数据挖掘，选择人工神经网络、决策树和支持向量机方法构造分类器，比较筛选出最适合的预测模型。结果显示：人工神经网络构建的预测模型准确度和性能较高，可达到89.11%。②任务期内模型预测整个任务期及任务期内的行为变量均具有预测功能，为了更好的辅助用户进行网络信息搜寻，任务期内的实时动态行为预测至关重要。输入特征向量选择：选择首次构建查询式时间、查询式关键词个数、唯一SERPs第一次停留平均时间、唯一页面第一次停留平均时间、唯一页面第一次停留平均时间/唯一SERPs第一次停留平均时间5个变量作为预测变量，尽管，上面5个变量是通过整个任务期计算而得，但是这些变量在任务搜寻过程中也可以计算，因此，这些变量可以作为任务期内预测变量。此外，经前分析，变量“唯一页面第一次停留平均时间”受任务执行者健康信息素养影响，应予以考虑剔除，但为避免输入变量过少，模型的可扩展性下降，本研究仍然将其纳入输入变量之中进行预测模型构建。分类器的构造方法及评价：本研究采用SPSS Clementine进行数据挖掘，选择人工神经网络、决策树和支持向量机方法构造分类器，比较筛选出最适合的预测模型。结果显示：人工神经网络构建的预测模型准确度和性能较高，可达到78.4%。（3）基于任务类型预测模型的健康信息个性化服务策略研究用户在完成不同信息搜寻任务时，需要网络系统提供的帮助亦存在差异。因此，挖掘不同任务类型下，用户信息搜寻行为模式，为个性化服务提供依据。只要系统监测到用户信息搜寻任务类型，就激发相应的辅助措施，可以帮助用户高效完成搜寻任务。①完善关键词推荐策略分析探讨不同任务类型下查询重构行为，可以帮助查询系统更好的提供查询提示，改善查询结果。101个调查对象共完成606个搜寻任务，其中222个任务包括查询重构行为，共计420次查询重构行为，以此为对象进行分析，结果显示：不同任务类型下，查询重构类型应用的频率、情境和有效性等均存在差异，系统可以据此提供有针对性的查询提示服务。②优化结果列表排序策略以单个查询为对象，划分为有效查询和无效查询，通过行为变量分析进行有效查询预测。101个调查对象共完成606个搜寻任务，搜寻过程中共键入1025个查询，其中有效查询742个，无效查询283个，将其作为分析对象，根据有效查询和无效查询的差异性，将查询过程变量和任务类型作为有效查询预测模型输入特征变量，采用人工神经网络方法构建模型，预测准确性为89.52%。为深入了解用户查询意图，优化结果列表排序提供可能性。【研究结论】（1）任务对网络健康信息搜寻行为具有重要影响。（2）根据网络健康信息搜寻行为能够预测任务类型且效果较好。（3）基于任务的健康信息个性化服务策略具有针对性。更多还原

【Abstract】 【Background&Objective】According to the WHO Research Report:1/3human disease can be avoided byhealth care prevention;1/3disease can be effectively controlled through earlydetection; therapeutic effect of1/3disease can be improved through informationcommunication. So, health information resources play an important role on healthcare, disease early intervention and assistant therapy. Internet has become animportant channel to obtain health information resources because of its massive anddiverse information and convenience. Although Internet is not an formal institutionsto get health information resources, it is an important supplement, which can supplyuser information window to in-depth understanding of health care, diseaseprevention and assistant therapy, and enhancement of health consciousness andmanagement. Some researchers find that it is beneficial to health behavior changeand health decision making ability enhancement when users seek health informationonline more often. In recent years, more and more users seek health information onWeb, but because of low health literacy level and Web personalized service, it is verynecessary to further improve the efficiency of seeking information and user satisfaction.Study on the user’s Web information seeking behavior and its characteristicsand regulations can provide users targeted help and improve the networkpersonalized service quality, which is very necessary and meaningful. Tasks are themain factors influencing the information seeking behavior, which is helpful tounderstand why people seek information, and how people select, acquire and useinformation. So, finding out the relationship of tasks and information seekingbehavior is the key point of solving the problem. Universal attention are paid to the influence of task and information seeking behavior, but how to make more targetedpersonalized service accordingly remains to be development. This study intends tofind out the influence relation of task type and information seeking behavior underthe network environment. Based on this, build task type predicting model. Inaddition, characteristics and regulations of information seeking behavior underdifferent task types are analyzed in order to stimulate the corresponding targeted help,which can provide new idea for improving Web personalized service.【Methods】This study adopts the method of experimental research; the data is collectedthrough questionnaires and screen video software.101undergraduate and graduatestudents are invited to participate this experiment.101valid questionnaires and606valid health information seeking task process are obtained.Data are analyzed bystatistical data analysis method. Pridicting model are built by data mining method.【Content&Results】(1) Relationship between Task Characteristics and Network HealthInformation Seeking Behavior①Build Task FrameworkInternal task basic characteristics include objectives, paths and outcomes, andthen extract task complex characteristics on the basic ones such as complexity,structural, etc. It is advantageous for definition and measurement of taskcharacteristics. According to task objectives characteristics in health field, task couldbe divided into three types in this paper, which are fact finding task, informationgathering task and decision making task.External task characteristics emphasis on performer characteristics includingprior knowledge and cognition characteristics, in addition, demographiccharacteristics are also analyzed. Prior knowledge includes health knowledge andnetwork knowledge, which is measured by health information literacy scale.Individual cognitive differences mainly reflected on the cognitive need and cognitivestyle dimension, respectively measured through cognitive need scale and cognitive style figure test. Demographic characteristics including gender, educational level areobtained through the questionnaire. In addition, the effects of website structure andcontent on information seeking behavior are analyzed.②Information Seeking Behavior VariablesInformation seeking behavior variables around seeking process include: queryformulation variables (Time on first query formulation, Query length, Query mode),result list valuation variables (Number of unique SERPs, Number of SERPs, Meanfirst dwell time on unique SERPs, Total dwell time on all SERPs, Mean dwell timeon all SERPs), the targeted site/web page operation variables (Number of savedpages, Number of content pages, Number of antique content pages, Mean first dwelltime on unique content pages, Total dwell time on content pages, Mean dwell timeon content pages, Mean first dwell time on unique content pages/Mean first dwelltime on unique SERPs, Total dwell time on content pages/Total dwell time on allSERPs), operation variables when results dissatisfied (Numbers of queryreformulation, Numbers of search engine or search scope change, Numbers of resultlist recheck), and task overall variables (Task completion time, Website sessions). Inaddition, subjective evaluation method is adopted to assess task performanceincluding degree of effort and satisfaction.③Relationship of Task Characteristics and Information Seeking BehaviorThe statistical analysis results show that:In task internal characteristics (objectives characteristics) aspect, there weresignificant differences among the three task type in regard to behavior variablesexcept inquiry mode and numbers of changing search engines.In task external characteristics aspect, there was significant differences betweenperformer cognition need in regard to total dwell time on SERPs, total dwell time oncontent pages. There were significant differences between performer healthinformation literacy in regard to total number of SERPs, mean first dwell time onunique content pages. There was a significant difference between performereducation degrees in regard to mean dwell time of content pages. There were not significant differences between performer gender and cognition style. In addition,clear site navigation, web links speed, web accessibility, site category had a greaterinfluence on information seeking behavior; The accuracy, authority, accessibility ofthe web content had a greater influence on information seeking behavior.(2) Task Type Prediction Model Research①Whole-session Prediction ModelFeature vector selection: according to the result of task effects on networkinformation seeking behavior variables, select suitable input variables. If there issignificant differences among task types in regard to the information seekingbehavior variable, then choose this variable into the model, in addition, if thevariable is influenced by other factors, then it is regarded as noise, remove it.Through the above analysis,14variables were chosen as follows: Task completiontime, Website sessions, Time on first query formulation, query length, Mean firstdwell time on unique SERPs, Number of unique SERPs, Mean dwell time on allSERPs, Number of saved pages, Number of content pages, Number of antiquecontent pages, Mean first dwell time on unique content pages/Mean first dwell timeon unique SERPs, Total dwell time on content pages/Total dwell time on all SERPs,Numbers of query reformulation, Numbers of result list recheck.The construction method and evaluation of the classifier: Data mining methodwas used to build prediction model with SPSS Clementine based on networkinformation seeking behavior variables. Artificial neural network, decision tree andsupport vector machine method were used to construct classifiers. The results showthat accuracyofartificial neural network model is higher, theaccuracycanreach89.11%.②Within-session Prediction ModelFeature vector selection:5information seeking behavior variables were chosenas follows: Time on first query formulation, Query length, Mean first dwell time onunique SERPs, Mean first dwell time on unique content pages, Mean first dwell timeon unique content pages/Mean first dwell time on unique SERPs. Although theabove5variables was calculated during the whole task session, but these variablescan also be calculated within task session, therefore, these variables can be used aswithin session predictor variables. In addition, according to the former analysis, the variable "Mean first dwell time on unique content pages" affected by performerhealth information literacy, should be removed, but in order to avoid poor modelextensibility because of few input feature vectors, this variable was still included.The construction method and evaluation of the classifier: Data mining methodwas used to build prediction model with SPSS Clementine based on networkinformation seeking behavior variables. Artificial neural network, decision tree andsupport vector machine method were used to construct classifiers. The results showthat accuracy of artificial neural network model is higher, the accuracy can reach78.4%.(3) NetworkPersonalizedServiceStrategyBasedonTaskTypePredictionModel①Perfect Keywords Recommended StrategyA total of420query reformulations under different task type were analyzed, theresults showed that there were differences among task types from queryreformulation type application frequency, situation and effectiveness, according towhich query hints can be provided.②Optimizing Results List Ordering Strategy1025individual queries were divided into742effective query and283ineffective query, and effective query can be predicted via information seekingbehavior variables, the accuracy of artificial neural network model can reach89.52%,which can be applied for further understanding user query intention.【Conclusions】(1)Task has important influence on network health information seekingbehavior.(2)It is feasible for predicting task type based on network health informationseeking behavior and accuracy of pridiction model is high.(3)Web personalized service strategies based on task are very meaningful.更多还原