By Dave DeFusco
Tirth Joshi, a student in the Katz School鈥檚 M.S. in Artificial Intelligence, won Best Poster Award at Network Science Informs AI 2026, a satellite event of the NetSci Conference 2026, for developing a new way to shrink large collections of data while doing a better job of keeping the features that show how the information is organized.
Joshi is addressing a basic problem created by the enormous amounts of information used in modern computing. The more data a computer must analyze, the more time and computing power the job can require. One way to make that work faster is to select a much smaller group of data points that can represent the larger collection.
鈥淢ost data analysis gets expensive as datasets grow,鈥 said Joshi. 鈥淎 landmark coreset is a small subset of the data chosen to stand in for the whole thing, so you can run the expensive computation on 250 points instead of 2,500 and still get an answer close to the truth.鈥
The challenge is deciding which points to keep. Existing approaches often try to spread the selected points throughout the data so that no part is too far from a representative. That provides good overall coverage, but Joshi found that it can miss places that are small but important, such as where two groups meet, where the shape of the data changes or where two regions are connected by a narrow bridge.
In his research, 鈥淧reserving Manifold Structure in Landmark Coresets: Morse-Seeded Sampling on kNN Graphs,鈥 Joshi compares the problem to deciding where to place weather stations. Putting stations evenly across a country would make sure that every location is reasonably close to one. But if scientists wanted to understand where weather conditions change, they might also want stations along coastlines or in mountain passes. Those places may cover little territory, but they can provide valuable information.
鈥淢y argument is that coverage alone treats the dataset as a cloud of points and ignores the fact that data usually has a shape,鈥 said Joshi. 鈥淭he places that define the shape, the boundaries between regions, the transitions, the thin connections, are exactly the places a coverage criterion has no reason to prioritize.鈥
Joshi鈥檚 method tries to find those important areas first. It looks at which data points are close to one another and identifies areas that appear to form meaningful regions or boundaries. It chooses some representative points from those locations and then uses an established method to spread the remaining representatives across the rest of the dataset.
Tests showed that the approach was especially effective on two datasets known as Swiss Roll and Two Moons, which researchers use because their shapes are easy to recognize. When Joshi reduced the datasets to one-tenth of their original size, his method did just as well as a widely used approach at representing the data as a whole. At the same time, it did a better job of keeping points from important areas where the data changes or different groups meet. The method was less successful with a more complicated dataset of handwritten numbers, showing that more work is needed before it can handle the complexity of real-world data as effectively.
The research could eventually help with tasks that depend on finding patterns in very large collections of information. Possible applications include creating simpler visual displays of complicated data, identifying groups within data and helping computer systems decide which examples are most useful to examine.
For Joshi, the project was also a personal milestone. He attended NSIA 2026 as the sole author of the research and said conversations with other researchers gave him new ideas about where to take the work next. The key result, he said, is that researchers may not have to choose between shrinking their data efficiently and preserving important features.
鈥淭he structural information is close to free rather than paid for,鈥 said. 鈥淚n other words, the experiments suggest it may be possible to keep a dataset small and manageable while still retaining more of the information that reveals its underlying shape.鈥
