Date of Degree

9-2018

Document Type

Dissertation

Degree Name

Ph.D.

Program

Physics

Advisor

Hernan Makse

Committee Members

Robert Haralick

Flaviano Morone

Gino Del Ferraro

Lucas Parra

Subject Categories

Statistical, Nonlinear, and Soft Matter Physics

Keywords

Network Science, Statistical Inference, Machine Learning, Socio-economy, Breast Cancer

Abstract

In this dissertation, we introduce the concept of network-based statistical inference methods of two types: network structure inference and variable inference. For network structure inference, we introduce correlation matrix, graphical Lasso, network clustering and identify the influencer in the network. For variable inference, we also introduce from Bayesian network, to Random Markov Field and Ising Model, Boltzmann and Restricted Boltzmann machine and the algorithm of Belief Propagation. Last but not the least, we introduce the most widely used neural network family and its two main types: Convolutional Neural Network and Recurrent Neural Network.

In Chapter 3 we provide an example of applying network structure inference algorithm to find the correlation between network metrics and socio-economic stats are introduced in this chapter. A mobile network with 108 nodes were constructed from mobile records to build the social networks by filtering the abnormal phone lines with a semi-supervised learning. Collect Influence (CI) is used as the proxy of network influence. A novel correlation (R2 = 0:95) is achieved by investigating the correlation between aggregated population which is based on both age and network metrics quantile. The result is validate by a marketing campaign.

In Chapter 4, we provide an example of combining large scale neural networks

to build a deep learning workflow to predict the pathology result of breast tumor based on MRI images. The work flow consist of three agents: Feature Extraction Agent which is a deep convolutional neural network transferred from inception v3. Image Selection Agent is a bi-directed recurrent neural network which evaluate the score of risk for each slice window and a Pathology Prediction Agent is to predict the pathology result based on the slices windows based on selection agent. The work flow is trained by reinforcement learning in order to automatically detect the location of tumor. Although the result indicates the workflow is able to capture the evidence of malignancy, the workflow still needs to be improve to increase stability.

This work is embargoed and will be available for download on Wednesday, September 30, 2020

Graduate Center users:
To read this work, log in to your GC ILL account and place a thesis request.

Non-GC Users:
See the GC’s lending policies to learn more.

Share

COinS