基于感知上下文的交互推荐算法研究

发布时间：2018-03-18 15:18

本文选题：上下文感知交互　切入点：推荐系统　出处：《山东师范大学》2017年硕士论文　论文类型：学位论文

【摘要】：伴随着数字化网络信息的普及,网络上的信息承载越来越大。推荐系统的作用就是让用户在海量的数据中快速得到自己需要的信息。各种各样的推荐系统也就应运而生,然而现在大多数针对推荐算法的研究都是关注提高推荐算法的精确度。然而我们熟悉的情境是,不同的用户的需求是不同的,现在的人们置身于不同的社会事务当中,每个人在不同的时间节点、地点、情境等会有不同的偏好。本文认为不同的用户所处的上下文信息不同,并且用户所处的上下文情境对推荐系统不可见的,这就要求我们的推荐系统去探知用户所处的上下文情境,并能够按照探知的上下文情境进行实时的推荐,但是现在的上下文推荐系统对于这种感知功能的关注度还是比较低的,像是感知交互的过程等,很多问题还需要做出深一步的研究。本文对上下文感知交互推荐进行的主要研究工作如下:1.提出一种上下文感知的交互推荐算法。针对实时的推荐系统,用户的上下文情境是处于多变的情况下的,很难直接用传统的用户偏好提取技术训练用户的偏好模型。并且针对这种上下文情境多变的情况,用户的上下文信息对系统是完全不可见的。本文认为针对用户上下文情境的探知,可以运用系统的推荐列表与用户对系统推荐的反馈列表进行交互的模式进行探知,本文提出一种交互模式,能够动态的探知上下文的改变,并在此过程中实时更新用户偏好模型,并用这种偏好模型监视可能出现的探知偏差。本文提出对于推荐列表的更新是一种动态随机分配最优化问题,本文提出TP-Learning算法,这个算法是对贪婪算法的一种改进,是基于启发式学习的,能够为用户得到一个效能最大化的推荐列表。本文提出的算法就是能够实时的去发现用户上下文的改变,并按照当前的上下文模式对用户偏好进行实时更新,并且运用这种机制提高推荐算法的表现。2.改进网络结构推荐算法(NBI)提出基于时间衰减和用户相似权重的二部图推荐算法(TUserCF)。基于二部图的推荐算法是将每个用户节点被赋予的资源值均分给相邻的节点,本文认为应该将用户选择物品的时间因素考虑进去;我们对于算法的改进是基于当前社交网络中社会化的因素,将用户之间的关系进行扩充。我们对于资源值的分配问题时,首先考虑的就是评分矩阵的影响,其次就是用户与用户之间的兴趣集合问题,好友集合参数,并且对于时间的衰减进行说明,并基于时间的衰减引入了分配系数的加权改变过程。最后按照最后的资源值进行推荐,但是这种资源值的推荐也会将评分的因素考虑在内。我们的算法显著提高了被推荐物品的准确性,能够使推荐更有效率,因此有很强的应用价值。3.本文设计了融合上下文信息的电子商务推荐系统框架,它是综合了用户具体属性、用户行为以及第三章中我们提出的上下文信息属性,构造N层笛卡尔积的属性集合,运用逻辑回归理论构建了融合上下文属性特征的电子商务网站推荐系统架构。这个架构是对全文推荐系统结构的一种总结,是综合了电子商务网站特点与推荐系统特点的综合产物。
[Abstract]:Along with the popularization of digital network information, network information carrying more and more. The recommendation system function is to allow users to quickly get the information they need in the vast amounts of data. All kinds of recommendation system also arises at the historic moment, but now most of the studies aimed at the recommended algorithm is the focus of the improved recommendation algorithm accuracy. However we are familiar with the situation is different, the user's needs are different, now people are living in a different social affairs, each place in the time node, different situation, will have different preferences. The context information of different users at different context and user's the recommendation system is not visible, this requires our recommendation system to ascertain the user's context, and to ascertain the real context according to the push Recommended, but attention to context recommender systems for this sensing function is still relatively low, as is the interactive process, many problems still need to make further research. Based on the context aware interaction recommended the main research work was as follows: 1. put forward a kind of interactive context aware recommendation algorithm. According to the real-time recommendation system, user context is in the changing circumstances, it is difficult to directly extract the user preference model for technical training. And the traditional user preferences in this context changing situation, the context information is completely invisible to the system. The discovery for the user context. And can use the recommendation list and user system interaction on the list of recommended feedback system model for discovery, this paper proposes a kind of interactive mode, can Dynamic discovery context changes, and updates the user preferences in the process model, and use this preference model to monitor possible deviation detection. In this paper, the recommendation list update is a dynamic random allocation optimization problem, this paper proposes a TP-Learning algorithm, this algorithm is a modified greedy algorithm, which is based on heuristic learning, can get a list of recommended maximum efficiency for users. This algorithm is capable of real-time to find user context changes, and real-time update of user preferences in accordance with the current context model, and the use of this mechanism to improve the performance of the.2. recommendation algorithm improved network structure recommendation algorithm (NBI) is proposed two similar weight decay time and user based recommendation algorithm (TUserCF). Two recommendation algorithm based on graph is each user node The point was given the resources value to adjacent nodes, we should take into account the goods time factors for our users; the algorithm is the social network of factors based on the relationship between users. We extend the problem of resources allocation for value, the first consideration is the scoring matrix the second is the impact between the user and the user's interest set, friends set parameters, and for the time attenuation is described, and the weighted change process distribution coefficient is introduced based on the time attenuation. Finally, according to the last resource values are recommended, but the resources of the recommended value will also score into account. Our algorithm significantly improves the accuracy of recommended items, can make the recommendation more efficiency, so it has strong application value of.3. is designed in this paper on the integration The electronic commerce recommendation system framework, it is a combination of a specific user attributes, context information of user behavior and attributes in the third chapter, we put forward the attribute structure of N layer of Cartesian product of a set of regression theory to construct the e-commerce website integration context attributes recommendation system architecture application logic. This architecture is a summary of recommendations the system structure of the whole thesis, is the integrative product of e-commerce website features and characteristics of the recommendation system.

【学位授予单位】：山东师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【参考文献】