上下文感知的移动用户新闻偏好获取及推荐算法研究

发布时间：2018-05-20 22:18

本文选题：移动新闻推荐 + 新闻偏好获取　；参考：《北京邮电大学》2017年博士论文

【摘要】：随着移动设备的普及和移动网络的迅速发展,越来越多的用户使用移动终端获取新闻等信息资讯。在这一现状下,如何根据移动新闻上下文感知的特点获取移动用户新闻偏好并进行移动新闻推荐,以提高推荐性能和移动用户满意度,成为上下文感知移动新闻推荐的主要研究任务。虽然传统的网络用户新闻偏好获取以及推荐已经在学术界和工业界取得了巨大的成功,但移动用户新闻偏好和移动新闻推荐通常受到多样性的,个性化的且动态变化的上下文影响,所以网络用户新闻偏好获取及推荐算法并不能直接应用于上下文感知的移动用户新闻偏好获取与推荐中。因此,本文针对移动环境中各种各样的上下文信息(如位置上下文,时间上下文和社会化上下文等),研究了上下文感知的移动用户新闻偏好获取及推荐等相关问题。本文的主要贡献有:1)提出了三个基于位置上下文的移动用户新闻偏好获取及移动新闻推荐算法。用户可以使用移动终端在任何地方浏览新闻信息,因此移动用户的新闻偏好通常都与位置上下文紧密相关。基于位置上下文的移动新闻推荐主要可以分为基于物理距离的方法和基于地理主题的方法。其中显式局部语义分析(Explicit Localized Semantic Analysis,ELSA)是最先进的基于地理主题的方法,它已经被证明要优于很多其他的主题模型,如显式语义分析(Explicit Semantic Analysis,ESA)和潜在狄利克雷分配(Latent Dirichlet Allocation,LDA)。然而,基于维基百科主题空间的ELSA由于受到高维度、稀疏性和冗余性等多种问题影响,大大降低了其新闻推荐性能。因此,为了克服上述问题,本文在ELSA基础上提出了三种地理主题特征模型,CLSA,ALSA和DLSA,分别将聚类、自编码和面向推荐的深度神经网络与ELSA集成,然后从维基百科主题空间抽取密集,抽象,低维度且非常有效的主题特征来描述新闻和地理位置。实验结果证明,在基于位置上下文的移动新闻推荐中,本文提出的三种算法的推荐性能均要优于最先进的地理主题模型ELSA,其中采用面向推荐的深度神经网络算法DLSA的改善效果最显著。特别地,由于本文提出的三种算法可以发掘用户在新位置上的潜在新闻偏好,因此它们也可以有效缓解“冷启动”问题。2)提出了两个位置感知的移动用户个性化新闻偏好获取及个性化移动新闻推荐算法。因为移动用户的新闻偏好通常与其位置上下文有关,所以许多研究工作已投入到位置感知的移动新闻推荐中向用户推荐离他们最近的新闻。然而,在现实情况中,移动用户的新闻偏好不仅仅依赖于他们的位置,还跟他们的个人兴趣紧密相关。因此,本文我们提出了一个基于显式语义分析的位置感知个性化新闻推荐算法(Location-aware Personalized News Recommendation with Explicit Semantic Analysis, LP-ESA),同时使用用户的个人兴趣和他们的位置上下文进行新闻推荐。然而,LP-ESA中基于维基百科的主题空间存在高维度,稀疏性和冗余性问题,大大降低了 LP-ESA的性能。为了解决这些问题,我们进一步提出了一个LP-DS A ( Location-aware Personalized News Recommendation with Deep Semantic Analysis )算法,利用面向推荐的深度神经网络来为用户、新闻和位置抽取密集的,抽象的,低维度且有效的描述特征。实验结果表明,LP-ESA和LP-DSA均显著优于基准方法。此外,与LP-ESA相比,LP-DSA可以在更短的时间里提供更有效的新闻推荐。3)提出了一个基于多维上下文的移动用户新闻偏好获取及移动新闻推荐算法。移动用户新闻偏好通常受到多种上下文的影响,如位置上下文、社会化上下文、时效性等。但是目前许多研究中使用的社会化上下文都是虚拟的,不能反映用户的真实社交关系,降低了新闻推荐性能。此外,大部分研究中新闻的时效性通过人为设定阈值进行时间过滤或时间建模来实现,但是这些方法没有考虑新闻时效性对移动用户新闻偏好的影响,而且人为设定阂值会为时效性计算带来一定误差。因此,为了解决上述问题,本文根据上下文感知的移动用户通信行为数据来推断其真实的好友关系,并利用新闻的点击流数据客观地分析获取新闻时效性,最后将移动用户社会化上下文,用户兴趣相似性和新闻时效性等多种上下文融合进行移动新闻推荐。实验结果证明,和现有方法相比,本文提出的方法显著提高了基于多维上下文的移动新闻推荐性能和推荐结果的时效性。4)提出了一个基于新颖性上下文的移动用户新闻偏好获取及移动新闻推荐算法。个性化推荐系统通常根据用户的历史数据获取用户偏好然后进行推荐。但是在新闻领域,新闻来源广泛,关于同一事件可能有多篇新闻报道,所以新闻候选集的冗余度比较高。在这一现状下,如果根据用户历史偏好进行新闻推荐,用户已浏览新闻会排在推荐列表的最前面。因此,新颖性检测对于实现高质量的个性化新闻推荐很重要。然而,现有的新闻新颖性检测方法大多基于几何距离或分布相似性,需要将当前新闻和用户历史数据中的新闻单独进行比较,时间成本很高,无法满足移动新闻推荐的实时性要求。因此,为了克服上述问题,本文利用LDA主题模型抽取新闻潜在主题,并将新闻看作样本而它们的潜在主题当作样本属性。然后将粗糙集和信息熵理论结合起来应用于新闻样本上获取属性权值,并计算每条新闻的总信息量,所以给定新闻的新颖性可以通过其与用户历史数据中新闻的总信息量之差的绝对值来快速衡量。接着我们进一步提出了一个正则化矩阵分解模型利用获取的新闻新颖性上下文和用户兴趣相似性来进行移动新闻推荐。实验结果表明,本文提出的算法提高了新颖性检测效率和基于新颖性上下文的移动新闻推荐性能以及推荐结果的新颖性。
[Abstract]:With the popularity of mobile devices and the rapid development of mobile networks, more and more users use mobile terminals to obtain information and information. In this situation, how to obtain mobile users' news preferences and recommend mobile news according to the characteristics of mobile news context awareness can improve the recommendation performance and mobile user satisfaction. While the traditional network user news preference acquisition and recommendation has been a great success in academia and industry, mobile user news preference and mobile news recommendation are often affected by diverse, personalized and dynamic contexts. Network user news preference acquisition and recommendation algorithms do not directly apply to context aware mobile user news preference acquisition and recommendation. Therefore, this paper studies context aware mobile user news for various contextual information (such as location context, time context and socialized context) in the mobile environment. The main contributions of this paper are: 1) the main contributions of this paper are as follows: 1) three mobile users' news preference acquisition and mobile news recommendation algorithm based on position context are proposed. Users can use mobile terminals to browse news information anywhere, so the new smell preference of mobile users is usually closely related to the location context. Mobile news recommendation based on location context can be divided into physical distance based methods and geographically based methods. Explicit local semantic analysis (Explicit Localized Semantic Analysis, ELSA) is the most advanced geographically based method, which has been proved to be superior to many other topic models, such as explicit Explicit Semantic Analysis (ESA) and potential Dirichlet distribution (Latent Dirichlet Allocation, LDA). However, ELSA based on the Wikipedia theme space is affected by a variety of problems, such as high dimension, sparsity and redundancy, greatly reducing its new recommendation performance. Therefore, in order to overcome the above problems, this paper is on the ELSA base. Three geographic feature models, CLSA, ALSA and DLSA, are proposed to integrate clustering, self coding and recommendation oriented deep neural networks with ELSA, and then extract dense, abstract, low dimension and very effective thematic features from the Wikipedia theme space to describe the news and geographic location. Experimental results show that the location is based on location. In the following mobile news recommendation, the recommendation performance of the three algorithms proposed in this paper is superior to the most advanced geographic topic model ELSA, in which the recommendation based depth neural network algorithm DLSA has the most remarkable improvement effect. In particular, the three algorithms proposed in this paper can discover the potential news preference of the user in the new location. Therefore, they can also effectively alleviate the "cold start" problem.2) and put forward two position aware mobile users' personalized news preference acquisition and personalized mobile news recommendation algorithm. Because the news preference of mobile users is usually related to the position context, many research workers have been put into position aware mobile news recommendation. However, in reality, the news preference of mobile users is not only dependent on their location, but also closely related to their personal interests. Therefore, we propose a position aware personalized news recommendation algorithm based on explicit semantic analysis (Location-aware Personalized). News Recommendation with Explicit Semantic Analysis, LP-ESA), using the user's personal interests and their location context for news recommendation. However, there is a high dimension, sparsity, and redundancy in the theme space based on Wikipedia in LP-ESA, which greatly reduces the performance of LP-ESA. In order to solve these problems, we enter One step proposed a LP-DS A (Location-aware Personalized News Recommendation with Deep Semantic Analysis) algorithm, which uses a recommended depth neural network for user, news and location extraction intensive, abstract, low dimensional and effective description features. Experimental results show that LP-ESA and LP-DSA are significantly better than benchmarks. In addition, compared with LP-ESA, LP-DSA can provide more effective news recommendation.3 in a shorter time. A mobile user news preference acquisition and mobile news recommendation algorithm based on multidimensional context is proposed. Mobile user news preference is usually influenced by various contexts, such as location context, socialized context, and time limitation. But at present, the socialized context used in many studies is virtual, which can not reflect the user's real social relations and reduce the performance of news recommendation. In addition, the timeliness of news in most of the studies is realized by time filtering or time modeling by artificial threshold, but these methods do not consider the news limitation. The effect of sex on the news preference of mobile users, and the artificial setting of the threshold will bring some error for the timeliness calculation. In order to solve the problem, this paper deduce the true friend relationship based on the context aware mobile user communication behavior data and analyze the news timeliness objectively using the News Click Stream Data. Finally, mobile news recommendation is carried out by a variety of context fusion, such as the socialization context of mobile users, user interest similarity and news timeliness, and the experimental results show that, compared with the existing methods, the proposed method significantly improves the performance of mobile news recommendation based on multidimensional context and the timeliness of the recommended results, which is based on the existing methods. Mobile user news preference acquisition and mobile news recommendation algorithm based on novelty context. Personalized recommendation system usually obtains user preferences based on user historical data and recommends. However, in the field of news, news sources are wide, and there may be multiple news reports on the same event, so the redundancy ratio of news candidate sets is compared. In this situation, if the news recommendation is based on the user's historical preference, the user has been browsing the news conference in front of the recommendation list. Therefore, novelty detection is important for the realization of high quality personalized news recommendation. However, the existing novelty detection methods are mostly based on geometric distance or distribution similarity. Comparing the news in the current news and the historical data of the user, the time cost is very high and can not meet the real-time requirement of the mobile news recommendation. Therefore, in order to overcome the above problems, this paper uses the LDA topic model to extract the news potential topics and regard the news as the sample and their potential topics as sample attributes. The theory of rough set and information entropy is combined to obtain attribute weights on news samples and calculate the total amount of information of each news, so the novelty of a given news can be quickly measured by the absolute value of the difference between the total information of the news in the user's historical data. Then we further propose a regularized matrix decomposition. The model uses the news novelty context and the user interest similarity to carry out the mobile news recommendation. The experimental results show that the proposed algorithm improves the novelty detection efficiency and the performance of the mobile news recommendation based on the novelty context and the novelty of the recommended results.
【学位授予单位】：北京邮电大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.3

【参考文献】