Modelling and Feature Selection with Applications to Big Data Problems

拟议研究的目标是继续(1)开发系统建模和特征选择技术，创建提取适用于短噪声实验记录和大数据问题的有用信息的方法，以及(2)将这些方法应用于物理和工业系统中的重要问题。拟议的研究将通过应用快速正交搜索（FOS）和改进正交搜索（MFOS）算法设计检测关键特征/实体和识别因果关系的方法。在相互作用的agent网络中，FOS和MFOS将逆向工程哪些关键网络实体控制其他网络实体的活动。FOS和MFOS每次选择一个特定的实体，并确定哪些其他实体是相互作用的代理。该方法将包括机器学习、深度学习、模式识别和分类，以确定最具影响力的实体和因果关系。该方法将FOS和MFOS集成到深度学习策略和模糊接口系统中。所提出的方法的一个独特方面是推导实体之间的抑制和激活的相互作用，尽管加密水平很高。FOS检测加密消息中的关键词，达到了所有测试方法中最高的准确率[mcgauhey等人，A Systematic Approach of Feature Selection for encrypted Network Traffic Classification， 2018年IEEE SysCon年会]。因此，改进以实现加密词的高速检测将是提出的方法的一部分。其他现实世界的问题包括理解心律失常。对于模糊模糊神经网络（FOS）、模糊模糊神经网络（MFOS）和并行级联识别（PCI）来说，关键是能够搜索非常大的候选集，从而快速找到预测某些输出变量值的最佳项。在可疑恐怖分子的网络中，我们可以确定哪些实体（个人或细胞）最能预测其他实体的时间活动（例如，使用通信设备，互联网时间等）。如果分配给小组成员的每个时间函数都是该人使用通信设备的时间，那么预测一个人的时间函数的候选术语可能不仅涉及其他小组成员的时间函数，还包括它们的交叉乘积。通过这种方式，其他人的活动变得显而易见，而无需表现出明显的互动。该方法的独特之处在于，它能够识别最不明显的候选对象——那些从未与其他实体进行明显通信，但却能最好地预测其他实体的网络活动的实体。FOS和PCI已成功用于基因调控网络的逆向工程[Wang Zhen，硕士论文，女王大学计算机学院，2010年10月]。然而，FOS、MFOS和PCI尚未应用于检测和破坏恐怖分子网络活动。我们将在并行实现中使用FOS和MFOS，这已被证明比快速傅里叶变换快10倍，快速傅里叶变换是物理学中在线相干成像的黄金标准。

Objectives of proposed research are to continue (1) to develop techniques for system modelling and feature selection, creating methods of extracting useful information applicable to both short noisy experimental records and big data problems, and (2) to apply such methods to important problems in physical and industrial systems. Proposed research will devise methods to detect key features/entities and identify causal relationships by applying Fast Orthogonal Search (FOS) and Modified FOS (MFOS) algorithms. In networks of interacting agents, FOS and MFOS will reverse engineer which key network entities control activities of the others. FOS and MFOS select one specific entity at a time and identify which others are interacting agents. The approach will include machine learning, deep learning, pattern recognition, and classification to determine both the most influential entities and cause-and-effect relationships. The approach will integrate FOS and MFOS into Deep Learning strategies and fuzzy interface systems. One unique aspect of the proposed methodology is deducing interactions of inhibition and activation between entities despite high encryption levels. FOS detects key words in encrypted messages, achieving highest accuracy of any method tested [McGaughey et al, A Systematic Approach of Feature Selection for Encrypted Network Traffic Classification, 2018 Annual IEEE SysCon]. Hence refinement to enable high-speed detection of encrypted words will be part of proposed methodology. Other real--world problems include understanding cardiac arrhythmia. Key to FOS, MFOS and parallel cascade identification (PCI) is ability to search very large candidate sets to rapidly find the best terms to predict the value of some output variable. In a network of suspected terrorists we can identify which entities (individuals or cells) best predict the time activities of other entities (e.g., use of communication devices, internet time, etc). If each time function assigned to a group member is the time that person is using a communication device, then candidate terms to predict one person's time function may involve not only the other group members' time functions but also cross--products thereof. This way other people's activities become apparent without ever demonstrating overt interaction. The proposed approach is unique in its ability to identify even the least obvious candidates - those entities never having apparent communication with others but yet best predicting the network activity of other entities. FOS and PCI have been used to successfully reverse engineer gene regulatory networks [Zhen Wang, MSc thesis, School of Computing, Queen's University, October 2010]. However, FOS, MFOS, and PCI have not been applied to detect and disrupt terrorist network activity. We will employ FOS and MFOS in parallel implementation which has proven to be up to 10 times faster than the Fast Fourier Transform, the gold standard in inline coherence imaging in physics.