Transformation Groups Seminar

Speaker: Sayantan Roy Chowdhury (Western)

"Toric regularity"

Time: 10:30

Room: MC 204

Algebraic Geometry

Speaker: Meagan James (Western)

"WAG: Elimination"

Time: 15:30

Room: MC 107

Geometry and Topology

Speaker: Rasul Shafikov (Western)

"Special embeddings in complex analysis"

Time: 15:30

Room: MC 107

Since Whitney's classical result of smooth embedding of manifolds into $\mathbb R^n$, one of the themes in geometry has been the existence of embeddings that preserve additional structure, for example, isometric embeddings of Riemannian manifolds or embeddings as symplectic or isotropic submanifolds into euclidean spaces with the standard Riemannian or symplectic structure. In this talk I will discuss some old and new problems concerning embeddings with special properties of complex or real manifolds into complex Euclidean spaces. We start with Stein manifolds and then discuss totally real embeddings and embeddings with special approximation properties. These problems have a topological component, such as Gromov's h-principle and Morse theory.

Colloquium

Speaker: Cristian Bravo Roman (Western)

"Are causal effects estimations the key to optimal recommendations under multi-treatment scenarios?"

Time: 15:30

Room: MC 107

Abstract: When making decisions that impact a specific context, it is essential to include a causal effect estimation analysis. This analysis allows us to compare potential outcomes under different treatment options or the control, aiding in selecting the best treatment option that leads to optimal results. However, merely estimating individual treatment effects may not suffice for making truly optimal decisions. To reveal this limitation, our study explores the incorporation of additional criteria, such as the uncertainty of these estimations, measured through a concept similar to the Conditional Value-at-Risk, commonly used in portfolio and insurance management. Additionally, we evaluate the inclusion of a specific prediction condition, particularly when the highest output is the desired outcome. With these intentions in mind, we propose a comprehensive methodology for multi-treatment selection, especially in situations where greater output is more desirable. Our suggested approach ensures the satisfaction of the overlap condition for comparing outcomes for treated and control groups. This involves training propensity score models as a preliminary step before employing traditional causal models — a crucial aspect often overlooked in other research. To illustrate the practical application of our methodology, we focus on the significant problem of credit card limit adjustment, which has historically been reliant on expert-driven choices. Analyzing historical data from a fintech company, we discovered that relying solely on counterfactual predictions is inadequate for generating appropriate credit line modifications. Instead, incorporating the two additional criteria significantly enhances the performance of the generated policy.

Short bio: Dr. Cristián Bravo is an Associate Professor and Canada Research Chair in Banking and Insurance Analytics at Western University, and Director of the Banking Analytics Lab. His research focuses on data science, analytics, and credit risk, particularly exploring multimodal deep learning, causal inference, and social network analysis to understand consumer-financial institution relations. With over 75 publications in prestigious journals and conferences, he contributes significantly to operational research, finance, and computer science. He frequently appears on CBC News’ Weekend Business Panel, and has been quoted by The Wall Street Journal, WIRED, CTV, The Toronto Star, The Globe and Mail, and Global News discussing topics in Banking, Finance, and Artificial Intelligence.

Graduate Seminar

Speaker: Esther Yartey (Western)

"Structural connectivity across datasets and species reveals community structure in cortex with specific connection features"

Time: 16:30

Room: MC 107

Advancements in neuroimaging technologies, particularly diffusion MRI, now allow reconstructing the long-range fiber connection patterns in the human brain. We study the network whose connection weights are determined by the number of fibers between individual brain regions. We study networks from the Human Connectome Project (HCP) and networks extracted from individual imaging subjects through a data processing pipeline developed in our group. By applying an algorithm to detect highly connected “communities” in these networks, we find a discrete set of communities appear robustly in the human brain. A specific community in the occipital lobe systematically displays high eigenvector centrality (EVC), a measure of the influence of nodes within a network. We explore the variations in these network structures among individuals and in retested subjects to isolate the sources of inter-individual variability. This result consistently appears across nearly all subjects and in a test-retest dataset. Similar community structure also appears in connectomes from macaque and marmoset brains, but the existence of an occipital lobe community with high EVC is specific to human connectomes. Taken together, these results reveal novel organization in the structural connectivity of the brain, derived from a fully data-driven approach, where clear community organization appears. This community organization relates to known functional divisions, such as visual and auditory sensory pathways, but also reveals community structure within higher-order areas, whose functional relevance can be studied in future work.