大规模PU学习

发布时间：2018-11-20 10:36

【摘要】：基于正样本和无标记样本的机器学习方法称为Positive and Unlabeled Learn-ing,简称 PU 学习。PU 学习在现实中有很多应用场景,例如某产品已有的用户数据可以视为正样本,企业须要基于这些正样本在海量未知用户中发掘潜在的新用户,该过程称为Lookalike扩展。本文根据应用场景的不同,将PU学习分为两类,一类是PU分类算法,一类是PU矩阵补全算法。第一类PU分类算法,主要针对某一特定的任务进行建模,例如某产品的Lookalike扩展。第二类PU矩阵补全算法,主要针对两类实体间的关系进行建模,例如单类协同过滤和多标记学习等应用。很多时候,除了能获取实体间的关系矩阵外,还有一些辅助的特征信息,例如单类协同过滤中,我们可获得用户或商品的特征数据,此时使用PU归纳式矩阵补全算法会取得更好的效果。已有的PU学习方法均是基于单机进行的,但是在大数据时代,数据规模不断增长,实用的机器学习算法应当具备可分布式的能力。本文在大数据处理平台Spark上,设计并实现了已有PU学习算法的分布式版本。此外,我们从多任务学习的方法中受到启发,提出了聚类PU归纳式矩阵补全的新方法。本文工作主要包括如下内容:一、实现了已有PU分类算法的分布式版本,包括分布式的两步骤方法和分布式代价敏感学习法,在企业级Lookalike数据集上,对比了各种策略的效果,并验证了算法具有一定的可扩展性。二、实现了已有PU归纳式矩阵补全算法的分布式版本,在推荐系统和多标记学习的标准数据集上进行了实验,验证了算法具有较强的可扩展性。三、提出了基于聚类思想的PU归纳式矩阵补全算法,实现了分布式版本,在推荐系统和多标记学习的标准数据集上,与已有的PU归纳式矩阵补全算法进行对比。新算法在AUC上获得较大提升,且具有一定的可扩展性。
[Abstract]:The machine learning method based on positive samples and unlabeled samples is called Positive and Unlabeled Learn-ing, for short PU learning. In reality, PU learning has many application scenarios, for example, the existing user data of a product can be regarded as positive samples. Based on these positive samples, enterprises need to discover potential new users among a large number of unknown users. This process is called Lookalike extension. In this paper, PU learning is divided into two categories according to the different application scenarios, one is the PU classification algorithm, the other is the PU matrix complement algorithm. The first kind of PU classification algorithm is mainly used to model a particular task, such as the Lookalike extension of a product. The second kind of PU matrix complement algorithm is mainly used to model the relationship between two kinds of entities, such as single-class cooperative filtering and multi-label learning. In many cases, in addition to obtaining relational matrices between entities, there are also auxiliary feature information, such as feature data for users or commodities in single-class collaborative filtering. In this case, PU inductive matrix complement algorithm will achieve better results. The existing PU learning methods are based on single machine, but in big data era, the data scale is increasing, and the practical machine learning algorithm should have the ability to distribute. In this paper, we design and implement the distributed version of PU learning algorithm on big data processing platform Spark. In addition, we propose a new method of clustering PU inductive matrix complement, which is inspired by the multitask learning method. The main contents of this paper are as follows: first, the distributed version of the existing PU classification algorithm is implemented, including the distributed two-step method and the distributed cost-sensitive learning method. On the enterprise level Lookalike data set, the effects of various strategies are compared. The extensibility of the algorithm is verified. Secondly, the distributed version of the existing PU inductive matrix complement algorithm is implemented. Experiments are carried out on the recommendation system and the standard data set of multi-label learning, and it is proved that the algorithm has strong extensibility. Thirdly, the PU inductive matrix complement algorithm based on clustering idea is proposed, and the distributed version is implemented. The algorithm is compared with the existing PU inductive matrix complement algorithm on the recommendation system and the standard data set of multi-label learning. The new algorithm improves greatly on AUC and has certain expansibility.
【学位授予单位】：南京大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP181

【相似文献】