联系我们

论文查询结果	返回搜索

论文编号：	9342
作者编号：	1120120810
上传时间：	2017/6/20 15:26:27
中文题目：	面向旅游吸引物POI的情境感知推荐方法研究
英文题目：	Methods Study for Context-Aware Recommendation of Points of Interest as Tourist Attractions
指导老师：	王芳
中文关键字：	情境感知推荐；POI；旅游吸引物；方法研究；用户偏好模型
英文关键字：	context-aware recommendation；POI；tourist attraction; method study; user preference model
中文摘要：	信息推荐在解决信息超载问题、提供个性化服务方面具有特有的优势，在众多领域均得到广泛应用。但传统信息推荐仅依靠“用户——项目”的二维关系并不足以解决问题。一些研究者提出将情境信息引入推荐过程，实现基于多维信息的推荐，情境感知推荐的概念由此而来。情境感知推荐是信息推荐的重要发展方向之一，近年来受到研究者越来越多的关注，但作为一个较新的研究领域，情境感知推荐的相关研究还需深入开展，特别是国内研究与国外研究还存在一定差距。同时，随着在线旅游信息搜寻活动的日益频繁，旅游信息推荐成为情境感知推荐重要的研究与应用领域之一。 基于以上背景，本研究将问题聚焦于旅游吸引物POI的情境感知推荐方法研究，面向特定的问题情境，解决情境感知推荐过程中的多个核心问题。本研究的主要工作包括： 第一，问题情境中影响用户信息需求和推荐任务的情境要素的识别与建模。本研究面向旅游吸引物POI的推荐，基于文献研究和用户需求文本的内容分析，对影响用户信息需求进而影响推荐任务的情境要素进行了发现与识别，构建了包括情境类型、情境属性和情境属性值的层次化情境模型，为推荐过程中引入何种情境要素提供了参考； 第二，用户偏好建模方法的设计与实验。该方法以POI描述文本的文本特征作为POI的属性特征，利用文本信息处理技术提取POI的基本特征，利用共词方法提取窗口共现词作为POI的补充特征，将POI表示为特征向量。在此基础上，改进了用户对POI属性的兴趣度计算方法，将用户偏好模型表示为兴趣度向量，利用余弦公式计算POI与用户偏好模型的相似度，基于相似度排序推荐。实验结果表明，该方法可以取得较好的推荐效果； 第三，单一情境信息兼容方法的设计。时空情境和天气情境的兼容采用了基于规则的方法，构建了过滤规则，将不符合特定约束的POI排除在推荐结果之外。对于层次化情境模型中其他情境属性的兼容，本研究构建了基于文本自动分类的情境信息兼容方法，将特定情境下的推荐问题转换为特定情境下的分类问题。本研究构建了情境化标注平台，收集了情境化标注数据，开展了分类实验，基于实验结果选取中心向量法为本研究的分类方法，最后利用所有标注数据训练生成19个情境属性值对应的中心向量，提出了基于余弦相似度的排序推荐方法； 第四，多元情境的整合推荐方法以及用户情境和用户偏好的整合推荐方法的设计与实验。本研究提出了情境预过滤和情境建模相结合的多元情境整合推荐方法，情境建模采用平均值法。用户情境和用户偏好的整合推荐采用了加权求和的方法，整合过程中引入用户社会情境作为不同赋权模式的条件。本研究设计了模拟实验，对本文提出的多个方法进行了有效性验证。最后构建了整合多元情境和用户偏好的情境感知推荐流程框架。 本研究的主要创新体现在：第一，面向情境感知旅游吸引物POI推荐的层次化情境模型的构建。已有研究较少关注情境要素的发现与识别问题，本研究则对问题情境中的情境要素进行了系统地发现与识别，在发现与识别的基础上构建了层次化情境模型；第二，提出了基于POI特征提取的用户偏好建模方法。这一方法中进行了两项具体的改进：一是将窗口共现词作为POI的补充特征，二是在用户兴趣度计算方法中引入用户评分极性、评分强度和评分阈值的概念，改进了兴趣度计算的方法；第三，提出了基于文本自动分类的单一情境信息兼容方法。情境信息兼容方法呈现多元化的趋势，本研究提出的基于文本自动分类的情境信息兼容方法是一种新的实现思路，在少量情境化标注数据的基础上即可以启动，是一种较为可行也可以取得较好效果的兼容方法；第四，提出了多元情境的整合推荐方法和用户情境和用户偏好的整合推荐方法。多元情境的整合推荐问题以及用户情境和用户偏好的整合推荐问题在已有研究中关注较少，本研究提出的方法是对已有研究的补充和发展。 在理论层面，本研究面向特定领域的情境感知推荐问题，构建应对情境感知推荐过程中各核心问题的方法体系，丰富和发展了情境感知推荐领域的研究主题、研究内容和研究成果。在方法层面，本研究主要关注方法的改进或发展，作为研究成果的方法可以为其他相关研究借鉴，同时，研究中收集数据、分析数据、设计方法的思路或方法，也可以从方法层面为其他研究提供参考。在实践层面，本研究的成果可以应用于多个相关领域，实现系统、服务或平台的建设与优化。
英文摘要：	Information recommendation has special advantages in dealing with the problem of information overload and personal information service, and is applied widely in many fields. Traditional information recommendation depends on the two dimensional relations between users and items, but sometimes this way of recommendation is not enough. Some researchers propose that additional information, such as context information, should be introduced into the process of recommendation and then recommendation on multi dimensional information could be provided. Thus the concept of context-aware recommendation is coined. Context-aware recommendation is one of the important developing trends of information recommendation, which is drawing more and more attention from researchers. But as a relatively new field, there are still lots of work need to do in the future. Meanwhile, the recommendation of tourism information is becoming one important research and application area of context-aware recommendation, with the development of online tourism information seeking activities. Under the above background, this dissertation focuses on the methods study of context-aware recommendation for POIs as tourist attractions, and tries to deal with the fundamental problems in the process of context-aware recommendation. The mean work of the study includes: Firstly, the study identifies and constructs the model of context factors that might influence users’ information needs and the recommending tasks. Based on literature study and content analysis of texts representing users’ information needs, the study identifies the influential context factors in the given area, and constructs the hierarchical context model which includes context types, context attributes and context values; Secondly, the study designs and evaluates one method for user modeling based on feature extraction of POIs. The proposed method takes text features of describing files as features of POIs, extracts basic features using text information processing technologies, takes co-occurring term pairs as complementing features, and represents POIs in the form of feature vectors. The method for calculating interest level of users on different features is also improved and thus user profiles are represented as interest level vectors. The similarity between POI vectors and user profile vectors is calculated using cosine function and the recommendation is created based on the ranking results of similarity calculation. The results of evaluating test show that the proposed method can get relatively good recommending effects; Thirdly, the study designs methods for integrating context information in the recommending process. The rule-based method is used for integrating time and space context and weather context, which sets filtering rules to filter out those POIs that couldn’t satisfy certain constrains. The study also designs one classification based method which converts the problem of contextual recommendation to the problem of classification under certain context. One contextual tagging platform is constructed and a cetain quantity of tagging data is collected. Based on this data set, one evaluating test is carried for choosing the best classification method and the results show Rocchio is the best one. Using all of the tagging data, 19 center vectors corresponding to the context values are created. The recommendation is then constructed based on the ranking results of cosine similarity; Fourthly, the study designs one method for integrating multi contexts and one method for integrating user context and user profile. Contextual pre-filtering and context modeling are used for integrating multi contexts and the averaging method is taken in context modeling. One weighted summation method is proposed for integrating user context and user profile and users’ social context is introduced into the process as the judgment condition for different weighting modes. A simulation experiment is carried for testing all the above methods. Finally, a workflow framework integrating multi contexts and user model is constructed. The mean innovations of this study include: firstly, constructing the hierarchical context model. Few studies focused on the problem of identifying context factors systematically, while this study focuses on this special problem and constructs the context model; secondly, proposing the user modeling method based on feature extraction of POIs. Two concrete improvements are proposed: one is using term pairs that co-occur in certain semantic window as complementing features of POIs; the other is improving the calculating method of users’ interest level based on the concepts of rating polarity, rating strength and rating threshold; thirdly, proposing the method for integrating context information based on text classification. There are multi methods for context information integration, while the method proposed is a new way which is more feasible in that it can run on a small collection of contextual tagging data and could get relatively good recommending results; fourthly, proposing the method for integrating multi contexts and the method for integrating user context and user profile. These two problems are less studied in the literatures and the methods proposed could improve and develop relative studies. Theoretically, the study tries to construct the methods system dealing with the fundamental problems in the process of context-aware recommendation in certain domain, which enriches and develops the topics, contents and results of existing studies. Methodologically, the study focuses on the improvement or development of methods, the methods as results will provide insights to other researches, meanwhile, the methods of collecting data, analyzing data and designing methods also have reference value. Practically, the results of the study could be applied in several fields for constructing or improving systems, services or platforms.
查看全文：	预览  下载(下载需要进行登录)