25–26 Jun 2026
''Vasil Levski'' National Military University
Europe/Sofia timezone

WEB Traffic Diagnostics by Techniques with Support Vector Machine, Naïve Bayes and Discriminant Analysis

Not scheduled
20m
''Vasil Levski'' National Military University

''Vasil Levski'' National Military University

Veliko Tarnovo, Bulgaria
Paper – Poster Presentation Information Technology

Speaker

Georgi Georgiev (Technical University of)

Description

Effective diagnostics of WEB services distribution is an important prerequisite for adequate traffic load service planning in information and communication infrastructures in serving corporate clients. Network administrators play a key role in specification and transmission medium metrics analysis in web traffic management. The paper proposes a Machine Learning approach for WEB traffic streams diagnostics and categorization of corporate business clients in urban areas. The approach envisages a pre-processing phase in network traffic analysis procedures, incorporating the following specific traffic metrics: Flows, IPv4 Packet size, IPv4 Packet size distribution, and Mean Transition Rate. The second stage of the research integrates analytical tools based on Support Vector Machine (SVM), Naïve Bayes (NB) algorithm, Linear Discriminant Analysis (LDA), and Quadratic Discriminant Analysis (QDA). The effectiveness of a system comprising C-SVM and Nu-SVM classifiers was evaluated using different sets of Kernel functions - Linear, Polynomial, Radial Basis Function (RBF), and Sigmoid types were applied. An advantage for Nu-SVM WEB traffic classification processes was established, where a maximum threshold recognition of 100.00% was achieved for the target classification groups defined. NB classifier training processes were conducted by Normal, Lognormal, Gamma and Poisson distributions applied to the input datasets. The highest classification accuracy of 100.00% was obtained for the Normal distribution, whereas when training the NB classifier with the Lognormal data distribution a significant reduction in this criterion below the 36.00% threshold was observed. Training and verification series of procedures were performed for Discriminant classifiers, respectively LDA, Diagonal Linear Discriminant Analysis (DLDA), Pseudo-Linear Discriminant Analysis (PLDA), QDA, Diagonal Quadratic Discriminant Analysis (DQDA) and Pseudo-Quadratic Discriminant Analysis (PQDA). In performance evaluation for the defined Linear and Quadratic types of DA classifiers, the highest recognition threshold of 90.00% was achieved in the WEB traffic streams analyzed in the case of PLDA classification procedures.

Authors

Mrs Ivelina Balabanova (Technical University of Gabrovo) Georgi Georgiev (Technical University of) Mrs Desislava Petrova (Technical University of Gabrovo)

Presentation materials

There are no materials yet.