FINDING PATTERNS IN A DIGITAL HAYSTACK [research]

Sriram Somanchi’s research approach uses machine learning to help companies increase performance metrics by discovering important trends in their data.

Sriram Somanchi is interested in anomalies. For the associate professor of business analytics at the Mendoza College of Business, it’s the blips in a vast ocean of data that have some truths to reveal.

“The main idea is that you have some anomaly — an indication that something is happening out there,” Somanchi said. “Each anomaly individually may not be interesting, but when you look at them together as a group, a pattern starts to emerge. And in the patterns, we start to find answers.”

Each day, companies receive and send several hundreds of gigabytes of data — roughly comparable to about 50 million plain-text emails or 5.2 million books — ranging from emails to video conferencing, uploaded photos and videos, and more complicated transactions such as processing thousands of purchases in just one hour.

Somanchi designs methodologies using machine learning that help companies sift through the thousands or millions of gigabytes of data to see which trends are emerging. It’s analogous to helping them find the microscopic needle in an immense haystack — the significant pattern or trend that will yield insights for managerial decision making.

Specifically, his research focuses on “subgroups” of data within massive datasets that tell a story researchers could use to determine their next course of action. He tackles complex problems by blending insights from social science with advanced machine learning techniques, creating new methods that connect these two fields.

Somanchi’s most recent research delves deeper into the intersection of machine learning, crowdsourcing and scientific discovery. His approach is helping shape an emerging and growing area of research in a fresh and impactful way.

The classic approach for business managers to make decisions that will improve some measure of performance is to define a change hypothesis and a set of metrics they want to measure. They then conduct an experiment to see which change gets better results; for example, showing some customers a website with a blue background and others one with a pink background to measure which one draws a better response.

Somanchi’s methodology of identifying subgroups provides a deeper analysis than the classic way of conducting a trial, which typically yields “average” results. His approach reflects the fact that there isn’t a single solution to a business problem that fits everyone and that “personalized” results are not only more helpful to companies but also more possible, given advances in artificial intelligence and machine learning.

“Average is good to begin with, but ‘average’ fits no one,” he said.

Somanchi, who joined Notre Dame in 2015, has a Ph.D. in information systems and management from Heinz College at Carnegie Mellon University. A graduate of the Machine Learning Department at CMU, he earned an M.E. in computer science from the Indian Institute of Science, Bengaluru, India.

Much of his recent research provides additional insights into how running experiments can affect the company’s findings in unexpected ways. Somanchi has worked with online retailers including eBay, which often conducts experiments to discover consumer preferences. The retailers track how many visitors to its app or website actually purchase products using metrics called conversion rates.

To improve these rates, the companies run experiments in two main ways: First, they can compare shopping behaviors across subgroups such as predefined age groups like Gen Z, Millennials, Gen X and Boomers. Alternatively, Somanchi explained, data analysts can identify natural customer groupings based on past purchasing patterns of existing customers.

In his recent study published in ACM Transactions on Information Systems, “Examining User Heterogeneity in Digital Experiments,” Somanchi analyzed real-world experiments spanning 1.76 billion sessions to identify subgroups based on user characteristics such as demographics, engagement, satisfaction and digital channels.

The product managers designed these experiments so that they could evaluate the impact of any change on their platform, such as a new search algorithm or recommendation system, a novel advertisement or changes to the user experience. The study was co-authored with Mendoza professors Ahmed Abbasi and Ken Kelley along with David Dobolyi of the University of Colorado's Leeds School of Business and Ted T. Yuan of eBay.

“Changes impact subgroups differently; what benefits one group may not help another, and reporting only the average can obscure important patterns,” Somanchi said. “We want to push experimentation platforms to create better tools that show these differences based on user characteristics. This would help team leaders, analysts and data experts make smarter decisions based on real evidence.

Another recent study, “Do Crowds Validate False Data? Systematic Distortion and Affective Polarization,” published in MIS Quarterly, challenged the common belief that “the wisdom of crowds” always leads to accurate information. The research was co-authored by Nicholas Berente, James H. Sweeny III and Alicia Sweeny Collegiate Professor of ITAO; Daniel A. Pienta of Haslam College of Business; Nishant Vishwamitra of the University of Texas at San Antonio; and Jason Bennett Thatcher of the University of Colorado-Boulder.

The authors investigated how emotional biases within subgroups can distort crowdsourced data, particularly when validating information about controversial events. The findings revealed that people’s political leanings and emotional reactions significantly affected how they validated information. More polarized individuals and those with stronger feelings of loyalty or betrayal to their group interpreted the same events differently, leading to distorted “ground truth” when their responses were combined. When dealing with politically charged or emotional topics, companies should account for these group biases rather than simply assuming that averaging many opinions will lead to accuracy.

His recent research analyzes trends in the health care industry, examining if either gender or racial inequities occurred during devices implanted among Medicare patients and determining the outcome of people admitted to the emergency room. Working with a health care startup in western Pennsylvania, he developed algorithms to ensure a smooth transition of patients in dynamic environments, such as ERs. Health care providers grapple with whether they have enough information about the patient to make a decision (e.g., discharge or admit in-patient from the ER) or collect additional information as the patient progresses through the care process (e.g., triage, vitals, labs).

“We developed the algorithm that learns the subgroup of patients who do not benefit from additional information, thereby sending early signals to aid clinicians and hospital administrators in their decision making,” he said.

Illustration by Carmona Errata. Photo by Michael Caterina/University of Notre Dame.

Sriram Somanchi is an associate professor of IT, Analytics, and Operations. His research harnesses the power of large-scale data and machine learning to discover subgroups that are statistically robust and theoretically grounded.

Published

“Examining User Heterogeneity in Digital Experiments”
ACM Transactions
Sriram Somanchi, Ahmed Abbasi, Ken Kelley, David Dobolyi and Ted Tao Yuan