Invitation to Research

Statistics Department Lightning Talks

September 2, 2020

Kris Sankaran

ksankaran@wisc.edu

Scientific Domains

• I’m curious about very small and very large ecosystems
• Microbiome: How should we describe all the bacterial ecosystems that surround (and inhabit) us?
• Climate Change: What can we learn about the large scale Earth Systems impacted by climate change?

Visualizing Data

• One of the best ways to empower a scientist is to streamline existing workflows
• We’re bad at understanding tables of numbers, but good at learning from figures
• Aim to build tools that are easy to learn and use

An R package showing interaction with U-Maps on cancer data. The left hand side are spatial imagery of a tumor, the right side arranges cells by genome expression. [link]

Investigating trends in opioid prescribing patterns. Selecting different regions of the PCA allow comparison of regions with very different patterns. [link]

Visualizing Models

• Modeling and visualization are complementary
• Both models and visualizations play a role compressing data into human-usable artifacts
• To recognize certain structure, it’s most efficient to work with both

Studying the structure of a regression tree applied to bicycle sharing data. The first split in the tree is between weekdays and weekends. [link]

Interactive error analysis on a glacier mapping model. The blue regions on the left are glaciers, they grey on the right are predictions. [link]

Latent Structure

• An alternative way to achieve compression is discover latent structure
• Groupings and gradients give concise stories behind complex data

Latent topics in a microbiome perturbation study, across different bacteria types. Some topics have more rapid recovery times than others. [link]

A simulation experiment about latent structure in spatial proteomic methods. Each square is a representative sample, and they’re arranged after learning latent features. [link]

Heterogeneity

• What should we do when we have several sensors into the same system?
• Proteomics + Microbiome
• Spatial + Expression
• What if our data are not I.I.D., but arranged hierarchically?
• Bacteria, across the phylogenetic tree
• Different glacier types, across many regions

A beautiful but impractical display of hierarchical structure in microbiome data. On the inside is a phylogenetic tree, the ring are time series associate with leaves. [link]

An illustration of integrative methods for microbiome and health data. You can read out species associated with different body types. [link]

Doubt

• Even after interacting with the data and models, there may be patterns we’re not sure about
• If you ran the experiment again…then what would happen?
• Providing diagnostics and measuring stability helps

Two descriptions of instability in feature learning algorithms. Points become clouds, and coefficient paths become bands [link].

Style

• Collaboration is important for gathering data about how science is done
• Clarifying the properties of popular workflows is useful,
• Validating (or debunking) folk wisdom
• Providing useful recommendations, either as methods or illustrations
• Research is a conversation