A Lassonde student’s paper was selected from more than 3,600 papers as a “best student paper finalist” at a major international conference hosted by the Institute of Electrical and Electronic Engineers (IEEE).
Saghar Bagheri, a master’s student, received the prestigious recognition for her work as lead author on the paper “Learning Sparse Graph Laplacian with K Eigenvector Prior Via Iterative GLASSO and Projection.” On June 10, she presented her research to delegates at the 2021 IEEE International Conference on Acoustics, Speech and Signal Processing, which was hosted virtually at the Metro Toronto Convention Centre.
The research outlined in the paper has the potential to offer improved modelling of data obtained from wireless sensors placed in geographically remote areas, such as forests, which could then be used to anticipate forest fires. Datasets derived from social network activity, political voting patterns and brain activity could also be modelled using the paper’s approach, to provide valuable insight to researchers in these areas.
While there is currently no shortage of big datasets, finding correlations within them and being able to model them remains a distinct challenge. Work currently being conducted in the research group of Gene Cheung, associate professor in the Department of Electrical Engineering and Computer Science at the Lassonde School of Engineering, is looking to examine graphs that represent networks of data points.
When most people hear the word “graph” they think of a bar graph or a line graph; however, to Bagheri, who is part of Cheung’s research group, a graph is a complex network with numerous nodes that represent data points. Pairs of data points are connected by edges, and together the nodes and edges create the network known as a graph. “A great example of a graph is someone’s social media network,” said Bagheri. “For example, people can be represented by ‘nodes’ and the connection between them can be represented by ‘edges.’ We can then study these graphs as sets of nodes and edges, and find patterns.”
These graphs are so massive and complex that it is impossible to extrapolate and model their behaviour without making some assumptions. Bagheri and her co-authors were able to make assumptions about the behaviour of the graphs by using some creative mathematics and approaching the problem in a way that no other research team has before.
“In order to model these massive networks, most researchers have primarily made assumptions within the nodal domain,” said Bagheri. “What we did was different – we made assumptions within the spectral domain.”
Bagheri said that using nodal space represents a direct approach, while using spectral space is an indirect approach to solving the problem of modelling datasets. The spectral space contains eigenvectors and eigenvalues. These eigenvectors can aggregate all the information from edges of a graph and can specify the most common patterns, which can then be used as fuel to construct a model that is significantly better at demonstrating correlations than existing algorithms. Bagheri and her co-authors demonstrated the effectiveness of their approach for image processing applications and believe it could be applied to any type of graph. In the future, they are interested in working with data mining applications and with brain data with the VISTA research group, of which Cheung is a core member.
Originally from Iran, Bagheri completed her undergraduate degree at Sharif University, studying math and computer science. It was her love of math that led her to pursue a graduate degree with Cheung’s research group. She defended her master’s thesis in June and is planning to stay on as a doctoral student in Cheung’s group.
“I always enjoyed studying math problems,” said Bagheri. “When I was looking for graduate supervisors, I learned that Dr. Cheung’s work was highly mathematical, which made joining his group an easy choice.”
Bagheri asserts that her work is far from done and that she wants to perform more experiments, try different spectral assumptions and work with the different datasets in the near future.