III: Small: Integrating Casual Discovery and Feature Selection with Streaming Features

随着图像处理、生物、金融等领域不断涌现的海量数据集的出现，传统的数据挖掘系统面临着新的挑战，即在动态的流特征环境中，随着时间的推移，新的特征不断涌入，挖掘知识和发现因果关系。这些挑战包括(1)特征量随着时间的推移不断增长，(2)巨大的特征空间，即使是未知的或无限大小的，以及(3)在学习开始之前并不是所有的特征都可用。这些挑战呼唤一种新的学习范式，具有不断增加的特征。在本项目中，我们将不断增加的特征量视为流特征，相应的学习问题称为流特征在线学习(OLSF)。由于现有的在线学习工作大多是处理不断增加的观察量但固定的特征维度的数据，OLSF提供了一个独特的机会来揭示和表征具有流特征的动态系统的模式趋势。本项目旨在解决OLSF的两个基本问题：(1)顺序增加特征维度的因果发现；(2)特征选择的因果关系。我们设计了OLSF中因果发现的新方法和算法，并在在线流特征学习的背景下，通过研究因果发现和特征选择之间的互惠关系，建立了它们之间的形式联系。为了对所提出的研究进行评估，我们在大量基准数据集上进行了经验研究，并在个性化新闻过滤和摘要中进行了广告特定的真实世界的案例研究，其中特征空间随着时间的推移而变化。该项目中的新算法和技术将提高我们使用有限资源的流特征从动态系统中发现知识的能力。该项目的方法谱不仅将丰富我们对动态系统的模式发现和机器学习的知识和理解，而且还将从流特征的角度为捕获和表征动态系统提供新的视角。

With the advent of emerging massive datasets in image processing,biology, finance, and so on, traditional data mining systemsface new challenges to induce knowledge and discover causalrelations in dynamic streaming feature environments, where newfeatures continuously stream in over time. These challenges include(1) continuous growth of feature volumes over time, (2) a huge featurespace, even of unknown or infinite size, and (3) not all featuresbeing available before learning begins. These challenges call for anew learning paradigm with continuously increasing features. In thisproject, we take the increasing feature volumes as streaming features,and the corresponding learning problem is referred to as OnlineLearning with Streaming Features (OLSF). Since existing onlinelearning efforts mostly deal with data with increasing observationsbut fixed feature dimensions, OLSF provides a unique chance to unfoldand characterize pattern trends for dynamic systems with streamingfeatures.This project aims to address two fundamental issues for OLSF: (1)causal discovery with sequentially increasing feature dimensions; and(2) causal relations for feature selection. We design novel methodsand algorithms for causal discovery in OLSF and establish formal connectionsbetween casual discovery and feature selection by investigating themutual benefits between them in the context of online stream featurelearning. To evaluate the proposed research, we conduct empiricalstudies on a large body of benchmark datasets, as well as with adomain-specific real-world case study in personalized news filteringand summarization where the feature space changes over time. Thenew algorithms and techniques in this project will advance our abilityto discover knowledge from dynamic systems using streaming featureswith bounded resources. The spectrum of the methods from the projectwill not only enrich our knowledge and understanding of patterndiscovery and machine learning for dynamic systems, but also provide anew view to capture and characterize dynamic systems from a streamingfeature perspective.