【作者】 张磊；

【导师】 陈俊亮；

【作者基本信息】 北京邮电大学 ， 计算机科学与技术， 2009， 博士

【摘要】 针对用户自身在实际需求,偏好特点和行为方式等方面的不同,个性化信息服务致力于满足用户个体的差异化信息需求。较传统的通用服务,个性化服务因为能够更好地表征、迎合用户的个性化偏好而受到了普遍的认可,个性化的相关技术也成为近年来一个新型的热门研究课题,受到了学术界和商业机构的广泛重视。本文围绕个性化技术中最为核心的两项,个性化推荐和个性化搜索中的若干关键问题进行研究、讨论,论文的主要工作包括以下内容：1.研究、探讨了协同推荐问题,在遵循基本协同的基础上,我们希望探寻、讨论新的有效推荐的研究思路。以此为基本出发点,本文提出了一种基于自低至高两个层面的多个BP神经网络进行项目评价预测的方法(Two-Level multiple Neural Networks-based Collaborative Filtering Recommendation Algorithm,简记为TMNN-CFRA)。两层面的多个BP神经网络协同工作,高层面BP网反向误差传播直至低层面多ANN进行网络权值修正,以此为基础借助用户评价等特征前向给出项目推荐预测。美国评测集Movielens上的实验评测验证了TMNN-CFRA算法的可行性和有效性。2.协作过滤推荐算法具有“冷启动”问题。“冷启动”问题的根源在于评价信息过于有限,推荐系统难以准确挖掘出用户偏好。本文提出了借助用户的模糊反馈信息改善冷启动推荐性能的基本研究思路(具体涉及2个算法)。对于项目推荐中棘手的冷启动问题可以从用户模糊反馈信息挖掘的角度展开研究,相对于完全地基于有限的项目评价本身的相似度测量改进等传统方法,这是一个相对比较新的研究基点,对于解决冷启动问题具有重要的意义。我们采用两个独立的算法研究、探讨了模糊反馈数据对于冷启动推荐的意义。其中,算法1采用后向传播的神经网络方法直接就模糊反馈数据本身进行学习,从“相对优劣”中挖掘用户对项目属性等的兴趣偏好；算法2对数据进行基础性变换,巧妙地从原本不具有可比性的模糊反馈数据和项目评价信息中抽取用户之间的相似度,以此为基础进行推荐预测。一般意义上而言,协作分析范畴的算法2较基于内容分析范畴的算法1具有更好的性能水平,初步验证了模糊反馈数据在冷启动阶段的积极意义。3.Web信息的爆炸式增长极大地激发了用户对于个性化的领域搜索服务的需求。本文提出并研究、实现了个性化的垂直搜索算法(Personazlied Vertical Search Algorithm,简记为PVSA),该算法以领域特征为出发点,借助领域主题偏好向量、领域元数据权重因子、检索名词差异化策略等4个策略有效挖掘、表征用户的领域个性化偏好,以此为基础生成基于用户偏好的垂直搜索算法,PVSA算法在个性化的领域搜索问题上取得了良好的效果。4.自动化的服务组合、服务推荐等是语义Web研究的重点。不同于完全地依赖本体进行服务推荐的思想,本文从统计学角度出发,提出了基于用户偏好的服务推荐算法(Preference-based Service Recommendation Algorithm,简记为PSRA),该算法首先基于Web服务语义进行无效后继服务过滤,然后基于职业本体、语义距离等针对人口统计学要素进行相似度计算,接下来融合人口特征至推荐评价,相对有效地给出综合人口统计学要素和评价信息的新的轻量的用户相似度度量,最后基于综合人口统计学要素和评价信息等特征的用户相似度给出满足用户个性化需求的后继推荐服务输出,PSRA在个性化服务推荐问题上取得了良好的效果。更多还原

【Abstract】 Personalized information services are focusing on the fulfillment of the personalized information demands of different users based on their preference characteristics, behivor patterns, etc. Comparing with the traditional ones, personalized services could effectively cater to users’ personal interests and correspondingly, they are widely accepted and becoming more and more popular. Lots of scholars and commercial organizations are paying their attentions to personalized services and many distinguished developments have been archieved in the past several years. In our paper, we present our research and discussion on two important techniques, the personalized recommendation and personalized search techniques. The main contributions are as follows:1. Focusing on the collaborative filtering process, we perform exploration and discussion for the new recommendation strategy. We present one novel method (Two-Level multiple Neural Networks-based Collaborative Filtering Recommendation Algorithm, TMNN-CFRA) for rating prediction in this paper. Multiple BP networks cooperating together, the higher layer neural networks propagates conversely the output deviation until to the lower layer neural networks to modify the network weights, and based on which, item recommendation prediction is accomplished by the forward process relying on the factors such as ratings, etc.. Experiment results on Movielens dataset show that TMNN-CFRA method is effective and feasible for item recommendation.2. Collaborative Filtering recommendation has cold-start problem. The root of the problem lies in that the ratings available are too limited, and recommendation system can not effectively mine users’ preferences with so scarce data. In our paper, we present the basic but novel idea to alleviate the cold-start problem by taking advantage of the mining of implicit feedback data (two strategies referred). Relative to the traditional cold-start improvement methods focusing completely on the sparse data, our idea has its significance. It presents an effective perspective to alleviate cold-start problem—fully mining by using corresponding algorithms rather than omitting the valuable implicit feedback data like the traditional methods. We present two independent strategies to exploit the significance of making use of users’implicit feedback for cold-start problem. In the first strategy, we use BP neural network to learn the feedback data itself, by which to mine users’prefences towards the factors such as item slot, etc., from the "relative superiority or inferiority". In the second strategy, we make the basic but effective transformation for the available data, and by which, the similarity information will be skillfully abstracted from the implicit feedback and item ratings which are of no comparability originally. In most cases, the second strategy belonging to collaborative filtering category will be more effective for item recommendation than the first one which belongs to the content-based analysis category and the significance of users’implicit feedback for cold-start recommendation has been preliminary demonstrated in our experiments.3. The rapid expansion of web information greatly stimulates the demands for personalized domain search services. In our paper, we present the personalized vertical search algorithm (PVSA). Based on domain characteristics, PVSA relies on four strategies including domain topic preference vector, domain metadata weight factors and distinguishing different weights of input terms, etc., to mine and present different domain preferences of different users. Consequently, personalized search outputs are obtained. Experimental results show that our algorithm holds the promise of effectively providering the personalized search capacity for different users.4. Automated service composition and service recommendation are essential for semantic web research. Not the same as the completely ontology-dependent idea for service recommendation, in our paper, we present preference-based service recommendation algorithm (PSRA) mainly from statistics perspective. Firstly, PSRA filters out the ineffective succeeding services based on service semantics, and then performs the demographic similarity calculation based on the strategies such as occupation ontology, semantics distance, etc.. In the following, by integrating demographic factors with recommendation ratings, PSRA effectively persents the new and light-weighted similarity measurement. Lastly, based on the redefined similarites between users and for the same current service, PSRA presents different succeeding recommended services to different users to meet their personalized needs. Experimental results show that our algorithm is feasible and effective.更多还原