基于属性提升与偏好集成的上下文感知推荐

发布时间：2018-07-26 09:34

【摘要】：上下文感知推荐研究近年来在推荐系统领域十分流行,因为通过对上下文信息的挖掘,这类推荐系统能更好地完成个性化推荐任务。上下文信息和推荐系统中的两个核心实体(用户和物品)有着紧密的联系,从这个角度看,上下文信息又可以被称为属性,主要包括用户属性和物品属性。围绕属性进行建模的技术不断发展且日新月异,然而,上下文推荐中的属性建模仍然面临这些问题:(1)现有的方法对属性的处理不够灵活;(2)已有的属性交互模型过于复杂而且没有针对性;(3)不同推荐任务对属性建模的要求不同;(4)通用的推荐算法受不同领域的属性影响较大;(5)不同记录对用户偏好的影响差异巨大。针对这些挑战,本文提出了属性提升与偏好集成两类策略来优化上下文感知推荐。首先,属性提升技术可以分别完成推荐系统中的评分预测和物品推荐两大任务;其次,偏好集成方法可以在属性提升的基础上改善预测结果;最后,偏好集成可以被改进成综合的推荐框架,完成一般化的推荐任务。具体的研究工作展开如下:(1)以评分预测为任务的局部提升与偏好集成研究以属性提升项为基础,局部提升技术通过用户、物品和属性类型三个角度完成属性交互。以梯度下降配合采样技术的局部学习策略,可以对提升框架进行高效训练,完成三个角度的偏好整合并形成局部的偏好预测值。以局部提升为基础的偏好集成方法使用了梯度提升树,实现多种偏好整合来得到用户的总体兴趣。局部偏好与总体兴趣一起,可以生成最终的评分预测值。实验证明,单纯的局部偏好预测已比流行的上下文方法(例如分解机)更准确,而且,在加入总体偏好之后,评分预测的准确度又有了进一步的提高。(2)以物品推荐为任务的全局提升与偏好集成研究全局提升的主要思路是通过属性邻居降低特定领域对属性建模的负面影响。首先,通过计算相似属性得到用户属性邻居和物品属性邻居;接着,对属性邻居进行单个集成,实现属性的领域无关化;最后,提出三种新型的交互方法完成全局物品推荐器。为了完善全局物品推荐器,偏好集成使用了局部低秩近似技术,给每个邻居赋予一个灵活的权重用来表达其贡献,进行多邻居集成与交互来完成推荐。实验验证了全局提升技术更不易受到特定领域的干扰,在加入全局偏好集成后,模型性能相对于先进的物品推荐器(如点对张量分解)有着更明显的优势。(3)以分部学习为基础的综合偏好集成框架与基于属性提升的两类偏好集成技术不同,综合的偏好集成方法把重点放在记录的分配与建模上。该框架的核心是分部学习策略,即通过记录划分、偏好挖掘、偏好集成三个有序的步骤完成推荐。首先,采用小组划分树完成对原始记录划分,并产生小组偏好;然后,以小组为单位,建立一个轻量级的回归模型用来捕捉组中特定用户的局部偏好;最后,通过对小组偏好和局部偏好的整合,得到用户的整体喜好,实现了综合的偏好集成推荐器。实验证明了通过灵活地调节模型粒度,综合偏好推荐器可以适应不同的数据集从而达到最佳性能,而且,该框架能以较高的准确度完成两大推荐任务。综上所述,通过提出属性提升与偏好集成策略,本文以多种方式优化了上下文感知推荐,完成了推荐系统中的两个主要任务。这是对属性建模与偏好预测的新尝试,为挖掘更深层次的用户兴趣提供了思路。因此,该主题有较大的理论研究价值与积极的实际应用意义。
[Abstract]:Context aware recommendation research has been very popular in the field of recommender systems in recent years, because this kind of recommendation system can better accomplish personalized recommendation tasks by mining context information. Context information and two core entities (users and items) in the recommendation system are closely linked. From this perspective, context information is the same. It can be called attribute, mainly including user properties and item attributes. The technology of modeling around attributes is constantly developing and changing day by day. However, the attribute modeling in context recommendation still faces these problems: (1) existing methods are not flexible enough to deal with attributes; (2) existing attribute interaction models are too complex and not targeted. (3) different recommendation tasks have different requirements for attribute modeling; (4) general recommendation algorithms are greatly influenced by the properties of different fields; (5) there is a great difference in the impact of different records on user preferences. In this paper, this paper proposes the two strategies of attribute promotion and preference integration to optimize context awareness recommendation. First, attribute lifting technology The two tasks in the recommendation system can be completed respectively. Secondly, the preference integration method can improve the prediction results on the basis of the attribute promotion; finally, the preference integration can be improved into a comprehensive recommendation framework to complete the general recommendation task. The specific research work is as follows: (1) the score prediction is used as the following. Local promotion and preference integration of tasks are based on attribute lifting. Local lifting technology completes attribute interaction through three angles of user, item and attribute type. Local learning strategy based on gradient descent and sampling technique can effectively train the lifting framework and complete the partial integration of the three angles and form a local. Preference prediction. The partial promotion based preference integration method uses a gradient lifting tree to achieve a variety of preference integration to get the overall interest of the user. Local preference, together with the overall interest, can generate a final score prediction value. Experiments show that the simple partial preference pretest is better than the popular context method (such as the decomposer). More accurately, and after adding the overall preference, the accuracy of the score prediction has been further improved. (2) the main idea of global promotion and preference integration for the task of the item recommendation is to reduce the negative impact of the attribute modeling on the specific domain through the attribute neighbor. First, the similarity property is calculated by calculating the similar attributes. The household attributes neighbors and property neighbors; then, the attribute neighbors are integrated individually to realize the domain independence of the attributes. Finally, three new interactive methods are proposed to complete the global item recommendation. In order to improve the global item recommendation, the preference integration uses the local low rank approximation technology to give each neighbor a flexible weight. It is used to express its contribution and carry out the recommendation of multi neighbor integration and interaction. The experiment verifies that the global lifting technique is less vulnerable to specific domain interference. After adding global preference integration, the performance of the model has a more obvious advantage than the advanced item recommendation (such as the point to tensor decomposition). (3) synthesis based on partial learning. The preference integration framework is different from the two type of preference integration technology based on attribute promotion. The integrated preference integration method focuses on the allocation and modeling of records. The core of this framework is the partial learning strategy, which is completed by the three orderly steps of recording division, preference mining and preference integration. The original record is divided and the group preference is generated. Then, a lightweight regression model is established to capture the local preference of a specific user in the group. Finally, the overall preference of the user is obtained through the integration of the group preference and local preference, and a comprehensive preference integration recommender is realized. The granularity of the model is adjusted alive, and the comprehensive preference recommender can adapt to different data sets to achieve the best performance. Moreover, the framework can complete two recommended tasks with higher accuracy. In summary, by proposing attribute lifting and preference integration strategies, this paper optimizes the context aware recommendation with a variety of formulas and completes the recommendation system. This is the two main task in the field. This is a new attempt at attribute modeling and preference prediction, which provides a way of thinking for deeper user interest. Therefore, this topic has great theoretical research value and positive practical significance.
【学位授予单位】：武汉大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.3

【相似文献】