基于监督学习的多样化推荐算法研究

发布时间：2018-09-07 06:51

【摘要】：伴随着大数据时代和移动互联网时代的来临,人们之间的交流更加频繁,关系更加紧密,信息匮乏的时代一去不返,信息过载的时代接踵而来。电子商务网站出现"井喷式爆发",人们的购物和生活方式发生了翻天覆地的变化。在这样的网络时代,传统的搜索算法不能为用户提供个性化的搜索列表,用户和市场的需求无法完全得到满足。因此,推荐系统作为一种个性化搜索工具应运而生,用于帮助用户对购物做出选择。推荐系统根据用户以往的查询历史、交易记录、用户资料或产品内容等信息,帮助用户对物品进行筛选、过滤,选择消费者最可能感兴趣的产品作为返回结果。这种方式一方面缩短了系统运行的时间,另一方面大大提高了用户获取信息的效率。当前的个性化推荐系统和相应算法更多地把目光投向于提高推荐系统的精确度,然而单纯考虑精确度的推荐算法往往将更多的流行物品推荐给用户,使得推荐列表变得单调,用户的视野变得狭窄,不能得到有价值的推荐信息。因此,推荐系统多样性作为推荐列表中物品丰富程度的代表,相关的研究越来越受到广泛的重视。提高推荐系统的多样性可以带来双赢的效果:一方面,依靠推荐系统的帮助,用户可以开拓视野,发现对自己有价值的商品;另一方面,商家企业可以利用推荐系统增加冷门商品的销售额并提升用户的满意程度,从而培养客户的忠诚度。现有的算法大多利用特定的多样性评价指标,采用启发式策略对物品进行重新排序,从而得到新的推荐列表。首先根据传统的基于精确度的算法选择一个用户偏爱的物品集合,然后通过最大化选定的多样性指标产生前k个物品的推荐列表。然而这些算法将推荐系统的精确度和多样性分成独立的两部分考虑,分别优化精确度和多样性两个目标函数造成问题求解效率下降,并不能很好的定义和解决监督学习问题。为了解决上述问题,本文利用监督学习方法在保证精确度的前提下研究推荐系统多样性的问题。其主要贡献有以下三个方面:(1)本文基于监督学习方法探究一个在保证精确度的前提下提高推荐系统多样性的算法。随后定义了一对耦合的优化问题:参数化矩阵分解问题和结构化支持向量机学习问题。(2)本文提出多样化协同过滤算法用以解决上述耦合问题,其中,结构化支持向量机学习得到一个推荐模型,用来生成每个用户的推荐列表。在模型学习的过程中,参数化矩阵分解算法贯穿在每一步优化过程中,用来生成最适合结构化支持向量机的用户和物品的潜在向量表示。(3)本文提出了新的基于集合的评价推荐系统精确度和多样性的测度:成对精确度和归一化主题覆盖多样性,并分别验证了新评价指标与基于排序的精确度和多样性评价指标之间的一致性。最后本文在不同数据集上进行大量重复的实验,通过分析实验结果来验证本文提出的算法在不同评价指标上的有效性并进行显著性检测。
[Abstract]:With the advent of the era of big data and mobile internet, people communicate more frequently, the relationship is closer, the era of lack of information is gone, the era of information overload comes one after another. In the Internet age, traditional search algorithms can not provide personalized search lists for users, and the needs of users and the market can not be fully met. Therefore, recommendation system as a personalized search tool came into being to help users make choices about shopping. Information such as material or product content helps users to screen, filter, and select the product that consumers are most likely to be interested in as a return result. This method on the one hand shortens the running time of the system, on the other hand greatly improves the efficiency of users to obtain information. Optical projection improves the accuracy of the recommendation system. However, recommendation algorithms that simply consider the accuracy often recommend more popular items to users, making the recommendation list monotonous, narrowing the user's vision, and unable to get valuable recommendation information. On the one hand, with the help of recommendation systems, users can broaden their horizons and discover products of value to themselves; on the other hand, businesses can use recommendation systems to increase sales of cold-door products and enhance their use. Most of the existing algorithms use specific diversity evaluation indicators and use heuristic strategies to reorder the items to get a new recommendation list. Firstly, a user-preferred item set is selected according to the traditional precision-based algorithm, and then the user-preferred item set is selected by maximizing the selection. However, these algorithms divide the accuracy and diversity of the recommendation system into two separate parts, and optimize the accuracy and diversity of the two objective functions respectively, resulting in the problem solving efficiency is reduced, which can not well define and solve the supervised learning problem. The main contributions of this paper are as follows: (1) Based on the supervised learning method, this paper explores an algorithm to improve the diversity of recommendation systems under the premise of ensuring accuracy. (2) In order to solve the coupling problem mentioned above, this paper proposes a diversified collaborative filtering algorithm, in which the structured support vector machine learns to get a recommendation model to generate a recommendation list for each user. (3) A new set-based measure for the accuracy and diversity of evaluation recommendation systems is proposed: pairwise accuracy and normalized topic coverage diversity, and the new evaluation index and ranking-based method are validated respectively. Finally, a large number of repeated experiments are carried out on different data sets to verify the effectiveness of the proposed algorithm in different evaluation indicators and to detect the significance of the results.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【相似文献】