Beyond Variance: Selecting Low-Rank Temporal Networks That Preserve Causality and Spreading

Imagine trying to understand the spread of a rumor, a virus, or a piece of information through a group of people. In the past, scientists often looked at this as a static map, a single snapshot showing who knows whom. But real life is not a photograph; it is a movie. People meet, talk, and part ways in a specific order. If Alice meets Bob, and later Bob meets Carol, Alice's message can reach Carol. If the order is reversed, the message stops dead. This sequence of events is crucial. To study these moving networks, researchers use a mathematical tool called a "movie" of connections, where every frame captures who is connected at that exact moment. The challenge is that these movies can be massive, containing thousands of frames and millions of potential connections. To make them manageable, scientists have developed ways to compress them, much like shrinking a video file, by keeping only the most important patterns and discarding the rest.

For a long time, the standard rule for deciding how much to compress was based on "variance," a statistical measure of how much the data changes from the average. The logic was simple: keep enough patterns to preserve 95% of the total change in the data, and throw away the rest. This works well if you just want to recreate the general shape of the network or see how often people meet on average. However, a new study by Madjid Eshaghi Gordji and Mohamadali Berahman at Semnan University in Iran asks a sharper question: does keeping 95% of the statistical change also keep the ability for things to actually travel through the network? They found that the answer is often no. A tiny, rare connection that appears only once might carry almost no statistical weight, yet it could be the single bridge that allows a virus to jump from one community to another. If a compression method discards that rare link to save space, the mathematical model might look almost perfect, but the story of how the virus spreads becomes completely wrong.

The researchers set out to find a better way to decide how much of the network to keep. Instead of just counting how much data is preserved, they tested whether the compressed version still allowed for "time-respecting paths." This means checking if a message could still travel from person A to person B in the correct order of time. They also tested how well the compressed network predicted the size of an outbreak, simulating a scenario where an infection spreads from a starting point to everyone it can reach. To do this, they analyzed five real-world contact networks, including data from a hospital ward, a workplace, a primary school, a high school, and a technology conference. These datasets recorded face-to-face interactions with high precision, capturing when people were close enough to potentially pass something along.

The team discovered that the standard method, which stops compressing once 95% of the variance is kept, often failed to preserve the network's ability to transmit information. In one specific test case they constructed, three patterns of connection captured nearly 99.9% of the statistical variance. The compressed network looked almost identical to the original when viewed as a whole. Yet, because it missed one rare bridge between two groups, the ability for a path to cross from one group to the other dropped to less than half of what it should have been. Adding just one more pattern, which contributed a tiny fraction of a percent to the total variance, instantly restored the full ability for the network to function correctly. This proved that a link can be statistically insignificant but causally vital.

When they applied this new, stricter test to the five real-world datasets, the results were consistent across most conditions. To ensure that the compressed network could still accurately predict how an infection would spread or how far a message could travel, they had to keep significantly more patterns than the standard method suggested in the vast majority of cases. In the hospital network, for instance, the standard method kept about 96 patterns, but the new method required 135. In the high school network, the standard kept 57, while the new method needed 78. Across the board, the "process-safe" rank—the number of patterns needed to keep the spreading dynamics accurate—was generally higher than the "variance" rank, holding true in 14 out of 15 different data and time-resolution combinations. The researchers found that the standard method often retained only about 36% to 58% of the total possible complexity of the network, whereas their new method retained between 46% and 78%.

The study does not claim that the old method is useless. If a researcher only wants to visualize the general structure of a network or see how often people meet on average, the variance-based compression is still a good tool. But if the goal is to understand how a disease spreads, how a rumor travels, or how a failure in a system might cascade, the old method can be dangerously misleading. It can make a network look safe and connected when, in reality, the critical pathways have been cut. The authors conclude that there is no single "best" number of patterns to keep for every purpose. Instead, the number of patterns must be chosen based on the specific question being asked. If the question is about movement and spreading, the compression must be tested against those specific movements, not just against the statistical energy of the data.

This work offers a clear lesson for anyone studying complex systems that change over time. The most important parts of a system are not always the loudest or the most frequent. Sometimes, the most critical element is a quiet, rare event that happens just once. By ignoring these rare events in the name of efficiency, we risk losing the very mechanism that makes the system work. The researchers have provided a new rule for scientists: do not just ask how much of the data you are keeping; ask whether the thing you are studying can still happen in the version you have left. In the end, the goal is not to make the data smaller, but to make sure the story it tells remains true.