Imagine a disease spreading through a network of connected people — like COVID-19 spreading through flights between cities. You have limited resources: you can only close a certain number of connections. Which ones do you cut? Does it matter if you choose carefully, or is picking randomly just as good?
I ran over 47,000 computer simulations across nine different network types to answer this. The key finding: it depends on the shape of the network and the timing of the intervention. When a network has distinct groups connected by a few critical bridges, cutting those bridges doesn't just slow the disease — it physically isolates regions so the disease can't reach them at all, creating safe zones.
Does it matter which connections you cut — or is random just as good?
When a disease is spreading, health officials have to make quick decisions: close which borders, quarantine which contacts, restrict which routes? One targeted strategy is to identify the connections that are structural "bridges" — the links that, if removed, would split the network into separate pieces. This is called betweenness centrality (BC) targeting. But past research disagreed on whether it actually works better than just cutting random connections.
This project set out to resolve that disagreement — and to figure out not just whether BC targeting works, but why, and under exactly what conditions.
BC wins by splitting the network into isolated islands — not by slowing the disease
Before this research, the leading explanation for why targeted removal works was that it makes the disease harder to spread overall — mathematically, it lowers the epidemic threshold. But the simulations showed something surprising: random removal actually lowers the epidemic threshold more than BC does — yet BC still wins.
The real reason BC works is different: it physically splits the network into disconnected pieces. Once a group of people (or computers, or cities) is cut off from the main network, the disease simply cannot reach them — no matter how contagious it is. These disconnected regions are "safe zones."
Targeted removal creates safe zones protecting 7–54× more people than random removal across real-world networks. The mechanism is isolation, not reduction in contagiousness.
Nine networks — from social media to biology
Not all networks look the same. A social network where a few celebrities connect millions of followers (hub structure) is very different from a network of tight-knit friend groups with a few cross-group connections (community structure). To test whether structure matters, I used four types of computer-generated networks that each isolate one structural feature — plus five real-world networks including a Facebook friendship graph, a power grid, and a biological neural network.
| Topology | Hubs | Clustering | Community | What BC Targets |
|---|---|---|---|---|
| BA Barabási-Albert | ✓ | ✗ | ✗ | Hub-hub edges (4.4× overrepresentation) |
| HK Holme-Kim | ✓ | ✓ | ✗ | Inter-neighborhood bridges |
| ER Erdős-Rényi | ✗ | ✗ | ✗ | No structural targets (null model) |
| LFR Benchmark | ✓ | ✓ | ✓ | Inter-community bridges |
BC outperformed random removal in 70% of real-world tests. The one exception — a neural network with almost no community structure — confirmed the rule: targeted removal only wins when the network has distinct groups to isolate.
The same strategy works through four different mechanisms
One of the most interesting findings is that BC doesn't win the same way every time. The network's shape determines how the targeted cuts do their work:
When a network is dominated by a few super-connected hubs (like major airports in a flight network), BC targets the connections between those hubs — making the whole disease less transmissible.
When the network has tight communities (like countries with strong internal travel), BC cuts the bridges between them — splitting the network into completely isolated safe zones with almost no impact on internal connectivity.
When communities are loosely defined, BC progressively carves the network into smaller and smaller pieces over time, gradually building containment zones.
Some networks have both hubs and community structure — BC works through both mechanisms simultaneously, making it especially powerful in these cases.
When you intervene matters just as much as how
A major and somewhat counterintuitive finding: the timing of intervention turned out to be the single biggest factor — more important than the network's shape and more important than which removal strategy is used.
- 01Early in an outbreak — when the disease is still spreading rapidly — targeted and random removal perform almost identically. Either approach gives modest benefit.
- 02Once the disease has settled into a stable endemic state (infecting roughly the same number of people each day), targeted removal becomes dramatically more effective — sometimes 20× more effective than random.
- 03The misleading early phase: In some networks, random removal actually appears to be winning in the first weeks of intervention. This reverses over time. A public health official who stops targeted intervention early because it "isn't working" would be making a mistake — the benefits accumulate slowly at first, then compound.
Proving it — not just observing it
Simulation results can show that something happens, but they don't explain why it has to happen. I also proved four mathematical theorems that give guarantees for community-structured networks. Think of these as the "why" behind the simulations:
In any community-structured network, the connections between communities mathematically must have higher betweenness centrality than connections inside communities. So BC is guaranteed to target the bridges before anything else.
BC can split a network into k isolated pieces by removing only a small fraction of its edges. Random removal would need to remove nearly all edges to achieve the same isolation.
BC's advantage holds until the disease reaches a specific prevalence level (determined by the network's structure). Beyond that point, the advantage disappears — the math predicts exactly when this crossover occurs.
The math shows that isolating communities (fragmentation) and making the disease harder to spread (threshold reduction) are driven by completely different structural properties of the network — which is why one strategy can do one without doing the other.
What this research established
- 01Network shape determines everything. The same targeted strategy produces dramatically different results depending on whether the network has hubs, communities, or neither. There is no universal "best" intervention.
- 02The textbook explanation is wrong. For decades, researchers assumed targeted removal works by lowering the epidemic threshold. It actually works by physically splitting the network into isolated pieces.
- 03Timing is the biggest variable. Intervening early during an active outbreak gives modest benefit regardless of strategy. Waiting until the disease stabilizes, then intervening with targeted removal, gives up to 20× greater effect.
- 04Targeted removal creates safe zones. Across real-world networks, BC-targeted removal protected 7 to 54 times more people than an equivalent number of random cuts — by isolating them completely rather than just slowing spread.
- 05The results are predictable in advance. Using just four measurable properties of a network, I built a model that can predict whether targeted removal will outperform random — before deploying any intervention.
Where these findings apply
The core question — which connections to cut, and when — shows up in surprisingly many real-world problems.
Pandemic border closures. During COVID-19, governments had to decide which international routes to restrict first. This research shows that in a world of tightly connected regional clusters (Europe, Southeast Asia), cutting the few inter-regional bridge routes first creates isolated safe regions far more effectively than closing a larger number of random routes.
Stopping malware outbreaks. When ransomware spreads across a corporate network, IT teams segment the network — cutting connections to contain the spread. The same framework applies: networks with departmental community structure benefit from bridge-targeted segmentation far more than random port blocking.
Misinformation on social media. False information spreads through bridge accounts — users who connect otherwise separate communities. Identifying and removing those few accounts fragments the spread network more effectively than removing a larger number of randomly selected accounts.