【作者】 高鹏；

【导师】 李武军；

【作者基本信息】 南京大学 ， 计算机科学与技术， 2017， 硕士

【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。更多还原

【Abstract】 Machine learning methods based on positive and unlabeled samples are called PU learning.PU learning is widely used in practical applications.For example,enterprise will discover new customers according to the existing customers which are considered as positive data,which is also called "Lookalike".PU learning can be divided into two categories according to different application scenarios.One category is called PU classification,the other is called PU matrix completion.The first one mainly builds model for a specific task,such as "Lookalike" for a product.The second one mainly builds model for the relationship of two sets of entities,such as one-class collaborative filtering and multi-label learning.In many cases,in addition to the relationship matrix between entities,there are some auxiliary feature information we can get,such as user or product features in one-class collaborative filtering,in this case we can use the PU inductive matrix completion algorithm to get a better result.Existing PU learning algorithms are all implemented on stand-alone machine.However,in big data era,practical machine learning algorithms should have the a-bility to be distributed.This article designs and implements the distributed version of existing PU learning methods on spark.In addition,inspired from multi-task learning,we propose a new model which is called cluster PU inductive matrix completion.This article includes the following three contributions:1.We implement the distributed PU classification algorithms,including distribut-ed two-step methods and distributed cost sensitive methods.Based on the big data set of Lookalike task,we compare all the methods.Moreover,these algorithms has a certain degree of scalability.2.We implement the distributed PU inductive matrix completion algorithm,and conduct experiments on the benchmark data sets of recommendation system and multi-label learning.We find that the algorithm’s scalability is very competitive.3.We propose a new method called cluster PU inductive matrix completion and design a distributed learning method for it.Based on the benchmark data sets of recom-mendation system and multi-tag learning,we compare our method with the state-of-art PU inductive matrix learning method.We find that our method has a better AUC,along with the competitive scalability.更多还原