Changing edges in graphical model algorithms

Graphical models are used to describe the interactions in structures, such as the nodes in decoding circuits, agents in small-world networks, and neurons in our brains. These structures are often not static and can change over time, resulting in removal of edges, extra nodes, or changes in weights of the links in the graphs. For example, wires in message-passing decoding circuits can be misconnected due to process variation in nanoscale manufacturing or circuit aging, the style of passes among soccer players can change based on the team's strategy, and the connections among neurons can be broken due to Alzheimer's disease. The effects of these changes in graphs can reveal useful information and inspire approaches to understand some challenging problems. In this work, we investigate the dynamic changes of edges in graphs and develop mathematical tools to analyze the effects of these changes by embedding the graphical models in two applications. The first half of the work is about the performance of message-passing LDPC decoders in the presence of permanently and transiently missing connections, which is equivalent to the removal of edges in the codes' graphical representation Tanner graphs. We prove concentration and convergence theorems that validate the use of density evolution performance analysis and conclude that arbitrarily small error probability is not possible for decoders with missing connections. However, we find suitably defined decoding thresholds for communication systems with binary erasure channels under peeling decoding, as well as binary symmetric channels under Gallager A and B decoding. We see that decoding is robust to missing wires, as decoding thresholds degrade smoothly. Surprisingly, we discovered the stochastic facilitation (SF) phenomenon in Gallager B decoders where having more missing connections helps improve the decoding thresholds under some conditions. The second half of the work is about the advantages of the semi-metric property of complex weighted networks. Nodes in graphs represent elements in systems and edges describe the level of interactions among the nodes. A semi-metric edge in a graph, which violates the triangle inequality, indicates that there is another latent relation between the pair of nodes connected by the edge. We show the equivalence between modelling a sporting event using a stochastic Markov chain and an algebraic diffusion process, and we also show that using the algebraic representation to calculate the stationary distribution of a network can preserve the graph's semi-metric property, which is lost in stochastic models. These semi-metric edges can be treated as redundancy and be pruned in the all-pairs shortest-path problems to accelerate computations, which can be applied to more complicated problems such as PageRank. We then further demonstrate the advantages of semi-metricity in graphs by showing that the percentage of semi-metric edges in the interaction graphs of two soccer teams changes linearly with the final score. Interestingly, these redundant edges can be interpreted as a measure of a team's tactics.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2017

Author dc:creator

• Chang, Linjia

Contributors dc:contributor

• Varshney, Lav R.

• graphical models
• decoding
• LDPC codes
• diffusion
• wiring errors
• data analysis
• semi-metric
• algebraic method

Statement dc:rights

• Copyright 2016 Linjia Chang

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/95404

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/95404