Dynamic ensemble selection for data streams and multi-view learning

多分类器系统(MCS)是机器学习和模式识别领域的研究热点。已经发表的几项研究表明，无论是从理论还是从经验的角度来看，它都比单独的分类器模型更具优势。最有前途的MCS方法之一是动态集成选择(DES)，它根据每个待分类的新样本动态地选择基分类器。在过去的几年里，DES已经成为多分类器系统文献中一个活跃的研究主题，因为最近的研究报道了动态集成比静态集成和单片分类器具有更好的性能。尽管DES方法最近取得了进展，但DES社区仍有大量未开发的领域。首先，大多数已进行的研究都考虑了一个固定的环境。然而，大多数现实世界的应用程序都有一个非静态环境，其中数据通常以数据流的形式出现，并且不断变化。处理非静态环境给DES方法带来了新的挑战，因为它们需要增量学习并处理数据流中发现的多个问题，例如添加新类、新功能和新的数据视图。其次，目前对DES技术的研究考虑了问题的单一视角(即单一特征空间)。然而，通常单视图数据不能正确地描述数据中的所有示例，因此需要多视图方法来提高泛化性能。文本分类、生物识别和物联网(IoT)等现实世界中的几个应用程序都从采用多视角学习方法中受益匪浅。因此，本研究项目的主要目标是提出DES方法来处理大数据流和多视角学习。这一目标将通过三个步骤来实现：(1)开发专门用于动态集合选择技术的集合生成方法；(2)适应动态集合选择以处理大量数据；以及(3)开发动态多视点集合选择方法。这一研究计划将产生稳健的动态集成模型，这些模型将有利于数据来自金融数据分类、假新闻检测和交通控制等流的应用。此外，我预计DES方法将对多视角学习领域产生重大影响，因为这种突破性的方法将使我们能够通过动态选择最相关的数据视角来有效地解决该领域的基本问题。

Multiple Classifier System (MCS) is an active area of research in machine learning and pattern recognition. Several studies have been published demonstrating its advantages over individual classifier models either from the theoretical or empirical points of view. One of the most promising MCS approaches is Dynamic Ensemble Selection (DES), in which the base classifiers are selected on the fly, according to each new sample to be classified. DES has become an active research topic in the multiple classifier systems literature in past years due to recent works reporting dynamic ensembles' superior performance over static ones and monolithic classifiers. Despite the recent advancement in DES methods, there are still vastly unexplored areas by the DES community. Firstly, most of the conducted research considers a stationary environment. However, the majority of real-world applications have a non-static environment where data usually comes in the form of data streams and is continuously changing. Handling non-static environments pose new challenges for DES methods since they need to learn incrementally and cope with multiple problems found in data streams such as the addition of new classes, new features, and new views of the data. Secondly, current research on DES techniques considers a single view of the problem (i.e., a single feature space). However, often a single-view data cannot properly describe all examples in the data, and a multi-view approach is therefore required to improve generalization performance. Several real-world applications, such as text classification, biometrics, and the Internet of things (IoT), greatly benefit from adopting a multi-view learning approach. Hence, this research program's main objective is to propose DES methodologies to deal with large data streams and multi-view learning. This objective will be handled through three steps: (i) Development of a pool generation approach specially crafted for dynamic ensemble selection techniques; (ii) Adaptation of dynamic ensemble selection for dealing with large volumes of data; and (iii) Development of a dynamic multi-view ensemble selection methodology. This research program will lead to robust dynamic ensemble models that will benefit applications where data comes inherently through streams such as financial data classification, fake news detection, and traffic control. Moreover, I expect DES methods to significantly impact the field of multi-view learning as this breakthrough methodology will allow us to efficiently solve fundamental problems in this field by selecting the most relevant views of the data on-the-fly.