Tahmineh Azizi, University of Wisconsin-Madison
Authors: Tahmineh Azizi
2023 AWM Research Symposium
Poster Presentation

Noise within data can significantly impact prediction of any study and can lead to poor results and reducing accuracy. Noise is anything that is not the "true" signal. Noisy data unnecessarily increases the amount of storage space required and can also adversely affect the results of any data mining analysis. It may have values close to your true signal. An outlier is something that is much different than the other values. In this study, we use different techniques to identify and handle different noise levels and their impact on big network data. We apply Mutual Information (MI) which is defined as a set of mathematical and statistical techniques with application in applied sciences, engineering and more specifically information communication. We perform Wavelet Leader Based Analysis (WLBA) which is robust under noise to be able to classify our data. In addition, we approximate the dimensional complexity of each network under different noise levels using Higuchi’s algorithm to obtain a complexity index for each network with different structure. According to our results based on algebraic topology and fractal geometry, with increasing the noise levels, the mutual information MI decreases. Moreover, with increasing noise level, Wavelet Leader Based Analysis (WLBA) method recognized a reduction in the spectrum width for all networks, while we noticed a wider range of spectrum when we impose smaller noise level on network structure. As maintained by Higuchi analysis, the dimensional complexity of noisy networks grows up with increasing noise level forced on network structure. These findings can be directly applied on big network data such as brain network, social network and other large scale network data.

Back to Search Research Symposium Abstracts