The New Jersey Institute of Technology's
Electronic Theses & Dissertations Project

Title:	Machine learning techniques for network analysis
Author:	Lateef, Irfan
View Online:	njit-etd2021-070 (x, 68 pages ~ 3.7 MB pdf)
Department:	Department of Electrical and Computer Engineering
Degree:	Doctor of Philosophy
Program:	Computer Engineering
Document Type:	Dissertation
Advisory Committee:	Akansu, Ali N. (Committee chair) Ansari, Nirwan (Committee member) Niver, Edip (Committee member) Khreishah, Abdallah (Committee member) Ulema, Mehmet (Committee member)
Date:	2021-12
Keywords:	Computer network management Eigenvalue and egienfunctions Machine learning algorithms Performance analysis Software defined networking Telecommunication networks
Availability:	Unrestricted
Abstract:	The network's size and the traffic on it are both increasing exponentially, making it difficult to look at its behavior holistically and address challenges by looking at link level behavior. It is possible that there are casual relationships between links of a network that are not directly connected and which may not be obvious to observe. The goal of this dissertation is to study and characterize the behavior of the entire network by using eigensubspace based techniques and apply them to network traffic engineering applications. A new method that uses the joint time-frequency interpretation of eigensubspace representation for network statistics as features for identification and tracking traffic flows based on the link level activity is proposed. Eigencoefficients (frequency domain feature set) and eigencomponents (time domain features) are jointly utilized to quantify their combined significance on the representation of each link data (each component of the link traffic vectors) in the eigensubspace. Several experiments are conducted using the joint time-frequency method to analyze the traffic data obtained from the Internet2 network. It is shown that the analysis with link-level resolution brings advantages for network traffic engineering applications. Specifically, this technique is applied to two scenarios: to identify large traffic flows and anomalous events in the network. Furthermore, machine learning methods are investigated to identify network paths using eigenanalysis of link statistics as the feature set. The merit of this method is validated by applying the technique on various network experiments. Eigenvectors and eigenflows in the subspace are jointly used as factors (features) for linear regression to forecast the network link traffic. It is demonstrated that the eigensubspace based autoregressive order two, AR (2), predictor is superior to the time-domain based predictor to forecast the link level traffic of a network. The unique contribution of this dissertation is using joint time-frequency interpretation of eigensubspace as features for identification of patterns and anomalies in conjunction with machine learning methods to automate the process and improve the accuracy of the method. This idea is not only applicable to the network analysis as demonstrated in this dissertation, but also applies to various fields of knowledge including medicine, finance and engineering. All of these fields have very large data sets in time domain, as well as complex patterns and relationships that exist among and are not discernible to human mind. This opens up a big area of application research using a combination of eigensubspace and machine learning. In the short term, the findings can be used to address 5G wireless energy optimization challenges wherein the problem involves a large number of communication channels serving an equally large number of users in time varying channel conditions. In the long term, the work can be expanded upon by using a Nonlinear autoregressive exogenous (NARX) machine learning model for forecasting in order to improve the accuracy, while also exploring other machine learning techniques such as Long short-term memory (LSTM) model.