Science Education Systems · Article 94. Maya, Jia Jun, Hana and Ethan are fictional learners used to make scientific reasoning visible. This article owns one distinct scientific job: network analysis—studying systems whose structure depends on relationships among entities. It does not replace systems thinking, spatial analysis or general graphing. Its job is to understand nodes, edges, connectivity, centrality, communities, diffusion, cascades and resilience.
The 50-second parent route
Some scientific systems cannot be understood by studying parts independently because the pattern of connections changes what the whole system can do.
The route is:
scientific question → entities → relationship definition → network representation → node and edge attributes → connectivity → centrality → community structure → paths → diffusion or flow → perturbation → resilience → bounded interpretation
The fastest diagnostic is to ask: If the same parts were connected differently, would the system behave differently? If yes, network structure is part of the Science.
This article extends How Scientific Systems Thinking Works, How Scientific Spatial Analysis Works, How Scientific Robustness Works and How Scientific Failure Analysis Works.
1. A network begins with nodes
Nodes are the entities being represented: organisms, people, proteins, neurons, stations, websites, species, components, institutions or concepts. The node definition determines what the network can answer.
2. A network also needs edges
Edges represent relationships: contact, friendship, flow, reaction, citation, transport, communication, dependency or another scientifically defined connection.
3. Maya’s first error is treating any co-occurrence as a connection
Two species are observed in the same park, so she draws an edge. Her repair is to define what the edge means scientifically: direct interaction, shared habitat or simple co-location are different.
4. Jia Jun’s first error is counting connections without direction
He treats “A sends information to B” as equivalent to “B sends information to A.” His repair is to determine whether the network is directed.
5. Hana’s first error is drawing a beautiful network and stopping
Her graph looks complex, but no scientific question is answered. Her repair is to link each network metric to a mechanism or decision.
6. Ethan’s first error is assuming the most connected node is always the most important
His repair is to distinguish several kinds of centrality because importance can mean direct reach, brokerage, global access or connection to influential neighbours.
7. Networks can be directed or undirected
Friendship may be treated as mutual in a simplified social network. Citation is directed. River flow is directed. Food-web predation has direction. Direction changes path and centrality calculations.
8. Networks can be weighted
An edge may represent one contact or one thousand contacts. Weight can represent strength, frequency, capacity, distance, probability or cost.
9. Weight meaning must be explicit
A larger weight can mean stronger connection or greater cost. Shortest-path algorithms interpret those cases differently.
10. Networks can be signed
Positive and negative edges can represent activation and inhibition, cooperation and conflict, attraction and repulsion.
11. Signed networks need careful interpretation
A node with many negative ties is not simply “highly connected” in the same sense as one with supportive ties.
12. Networks can be temporal
Connections form and disappear. Contacts change by hour, ecological interactions by season, and infrastructure topology after failures.
13. Static snapshots can hide network dynamics
Two networks with the same total number of edges can behave differently if their connections occur at different times.
14. Multiplex networks contain several relationship types
The same entities may interact through communication, finance, transport and social ties simultaneously. One layer can reinforce or compensate for another.
15. Bipartite networks connect two node types
Plants to pollinators.
students to courses.
authors to papers.
genes to diseases.
Edges occur between node classes rather than within one class.
16. Projecting bipartite networks can create misleading density
Two plants share an edge in a projected network because the same pollinator visits them. That is an inferred relation, not a direct interaction.
17. Degree is the simplest local connectivity measure
Degree counts edges incident on a node. In directed networks, indegree and outdegree separate incoming and outgoing connections.
18. High degree means local reach, not universal importance
A node can have many connections inside one small cluster while being unimportant for connecting the wider network.
19. Degree centrality normalises degree
It expresses direct connectivity relative to the number of possible neighbours, making some comparisons across network sizes easier.
20. Betweenness centrality measures brokerage
A node has high betweenness when many shortest paths between other nodes pass through it. Such nodes can act as bridges or bottlenecks.
21. A low-degree bridge can have high betweenness
A node connecting two dense communities may have only two edges yet be structurally critical.
22. Closeness centrality measures global access
A node is central when it can reach all others through relatively short paths.
23. Closeness needs care in disconnected networks
If some nodes cannot be reached, ordinary definitions break or require modified forms.
24. Eigenvector centrality values influential neighbours
Connections to well-connected nodes count more than connections to isolated nodes.
25. PageRank-like ideas add direction and flow
Importance can depend on incoming links weighted by the importance of the linking nodes, with adjustments for random movement or damping.
26. Centrality measures answer different questions
Degree: who has many direct ties?
Betweenness: who bridges?
Closeness: who can reach others efficiently?
Eigenvector: who connects to influential nodes?
27. No centrality is universally “best”
The appropriate metric follows the mechanism. Information spreading, epidemic exposure and infrastructure load may require different concepts of importance.
28. Primary Science can learn network thinking through simple connection maps
Food chains can become food webs.
Electrical circuits form connected pathways.
Water networks connect reservoirs and pipes.
The child learns that relationships create structure.
29. Primary 3 can count direct connections
Draw five nodes and edges. Ask which node has the most direct neighbours and whether that makes it the best bridge.
30. Primary 4 can compare paths
How many steps from A to D? What happens if one connection disappears? Path thinking introduces network resilience.
31. Primary 5 can identify bottlenecks
Find the one bridge connecting two groups. Predict what happens if it fails.
32. Primary 6 can distinguish local from global importance
A node with many neighbours inside one group may matter less to the whole network than a node connecting groups.
33. Secondary Science can formalise adjacency matrices
A network can be encoded as a matrix where rows and columns represent nodes and entries represent edges or weights.
34. Adjacency matrices connect graph theory to linear algebra
Paths, centrality and diffusion can be studied through matrix operations and eigenvalues in more advanced analysis.
35. Edge lists are another representation
Each row contains source, target and perhaps weight or timestamp. Edge lists are efficient for sparse networks.
36. Representation should preserve meaning
Converting a weighted directed network into an unweighted undirected network can remove the mechanism being studied.
37. Paths connect nodes through sequences of edges
A path describes a possible route for information, infection, material or influence under the network interpretation.
38. Shortest paths depend on the cost definition
Fewest edges, least travel time, lowest resistance and greatest capacity are different path problems.
39. Geodesic distance is network distance
It is the length of a shortest path under the chosen edge metric.
40. Diameter measures the longest shortest path
It describes how far apart the most distant connected nodes are in path terms.
41. Average path length describes typical separation
Short average paths can support rapid spread or communication.
42. Disconnected components are separate sub-networks
If no path connects two groups, diffusion cannot cross without a new edge under the model.
43. The giant component matters in large networks
A network can contain one large connected component plus many small fragments. Connectivity of the giant component affects system-wide reach.
44. Density measures how many possible edges exist
A dense network has many realised connections relative to possible ones. Density changes with network size, so comparisons require care.
45. Clustering coefficient measures closed neighbourhoods
If A connects to B and C, are B and C also connected? High clustering indicates locally cohesive groups.
46. Triangles can signal redundancy or social closure
In infrastructure, triangles may offer alternate routes. In social networks, they can indicate tightly connected groups.
47. Communities are groups with denser internal connections
Community detection seeks modules, clusters or compartments in the network.
48. Communities can reveal functional organisation
Protein interaction modules, social groups, ecological guilds or subnetworks can correspond to distinct functions.
49. Community detection is not one objective truth
Different algorithms and resolution settings can produce different partitions. Community structure should be interpreted with domain evidence.
50. Modularity is one community-quality measure
It compares within-community edge density with a reference expectation. High modularity can indicate stronger division into modules.
51. Resolution limits can hide small communities
Some methods merge meaningful small groups because optimising a global criterion favours larger modules.
52. Hierarchical communities exist
Small modules can sit inside larger modules, like cells inside organs or local teams inside organisations.
53. Network motifs are recurring small patterns
Feed-forward loops, triangles and other subgraphs can occur more often than expected and suggest functional building blocks.
54. Motif frequency depends on the reference model
A pattern is only “overrepresented” relative to an appropriate null network.
55. Degree distributions describe connectivity heterogeneity
Some networks have relatively uniform degrees; others contain a few hubs with many connections.
56. Heavy-tailed degree distributions create hubs
A small number of highly connected nodes can dominate pathways, spreading and vulnerability.
57. “Scale-free” should not be used casually
Real networks may show heavy tails without following a pure power law. Claims need formal comparison rather than log-log visual impression alone.
58. Preferential attachment is one hub-generating mechanism
New nodes are more likely to connect to already well-connected nodes, creating cumulative advantage.
59. Homophily creates another network pattern
Similar nodes preferentially connect: same interests, traits, habitat or function. Homophily can create clusters without direct influence.
60. Influence and selection are easy to confuse
Connected people behave similarly. Did they influence each other, or connect because they were already similar? Network causality requires design beyond association.
61. Network diffusion models spreading
Information, infection, innovation or failure can move along edges under rules of transmission.
62. Diffusion speed depends on topology
Short paths and hubs can accelerate spread; modularity can slow global spread while allowing fast local spread.
63. Threshold models require enough local exposure
A node changes state only after a fraction or number of neighbours have changed. Social adoption and cascading behaviour can follow threshold-like rules.
64. Simple contagion and complex contagion differ
One exposure may be sufficient for infection-like spread. Behavioural adoption may require reinforcement from several neighbours.
65. Epidemic models on networks depend on contact structure
Who contacts whom changes transmission opportunities. Average contact rate alone can hide superspreading hubs or isolated groups.
66. Network interventions can target hubs
Removing or protecting high-degree nodes may reduce spread efficiently under some mechanisms.
67. Betweenness-targeted interventions can protect bridges
Nodes connecting communities may be strategic points for containment even if their degree is modest.
68. Targeted strategies depend on accurate network data
Missing edges or delayed contact information can identify the wrong “critical” nodes.
69. Cascading failure is a network phenomenon
One component fails, load redistributes, neighbours exceed capacity, and failure spreads.
70. Robustness asks how connectivity survives perturbation
Remove nodes or edges and measure component size, path length, reachability or service level.
71. Random failure and targeted attack differ
A hub-dominated network may tolerate random node loss yet fragment quickly when hubs fail.
72. Resilience includes recovery
Can the network restore connections, reroute flow or reconfigure after damage? Static robustness is only part of resilience.
73. Redundancy creates alternate paths
Multiple routes can maintain flow if one edge fails. But redundant edges sharing the same physical corridor may share the same hazard.
74. Common-cause failure remains possible
Two “independent” network routes may share power, geography or software. Topological redundancy is not always functional independence.
75. Flow networks add capacity
Roads, pipes, power grids and communication systems have limits on how much can pass through edges and nodes.
76. Maximum-flow ideas reveal bottlenecks
The total throughput between source and sink is limited by critical cuts under the model.
77. A min-cut identifies vulnerable separation sets
A small set of edges may disconnect two important regions. Those edges can be high-value resilience targets.
78. Flow direction can change with failure
Rerouting after disruption can overload alternate paths. Dynamic load matters.
79. Worked case: school friendship network
Students are nodes; regular study-partner relationships are edges. Degree shows who has many partners, betweenness can reveal bridges between groups, and communities may represent peer clusters.
80. The school case teaches privacy
Real social-network data can expose sensitive relationships. Ethical handling, consent and aggregation may be necessary.
81. The school case teaches selection versus influence
High-performing students cluster. Are they improving one another, or choosing peers with similar habits? The network alone cannot decide.
82. Worked case: food web
Species are nodes; feeding relationships are directed edges. Removing one species can affect predators, prey and indirect competitors.
83. The food-web case teaches cascading effects
One edge change can alter several populations because effects propagate through pathways.
84. The food-web case teaches weighted edges
Rare feeding and dominant energy pathways should not always count equally.
85. Worked case: transport network
Stations are nodes; rail links are edges. Degree measures direct connectivity, betweenness can reveal interchange importance, and travel-time weights define shortest paths more realistically than edge counts.
86. The transport case teaches resilience
Close one station and recalculate reachable routes. Some failures create minor detours; others fragment the network.
87. Worked case: concept network in Science learning
Concepts are nodes; prerequisite or explanatory links are edges. A learner may know many isolated facts but lack bridges connecting energy, particles, forces and systems.
88. The learning-network case teaches bottlenecks
One missing bridge concept can block transfer across several topics. Repairing that node or edge can produce disproportionate improvement.
89. Primary 3: build a simple network
Use five organisms in a food web. Count direct links and trace one path of influence.
90. Primary 4: remove one node
Ask what relationships disappear and which nodes become isolated.
91. Primary 5: compare local and global importance
Identify the node with most links and the node joining two groups. Explain why they represent different kinds of importance.
92. Primary 6: trace a cascade
Change one species, station or component and follow several downstream effects.
93. Secondary Science: add centrality and community structure
Students can calculate simple centralities, compare communities, analyse paths and model spreading.
94. Secondary Science: add weighted and dynamic networks
Edges can have capacity, strength, time and direction. The network becomes closer to real systems.
95. Network analysis and systems thinking are different
Systems thinking is broader: parts, feedback, boundaries, emergence. Network analysis quantifies relationship structure explicitly.
96. Network analysis and spatial analysis are different
Spatial analysis uses physical or geographic relationships. Network analysis can involve non-spatial connections such as citation, molecular binding or logical dependency.
97. Networks can still be spatial
Road, river and power networks combine graph structure with geographic location.
98. Network analysis and mechanism are connected
A path suggests a possible route of influence. Mechanistic evidence is still needed to show that the influence actually travels along it.
99. Network analysis and failure analysis are connected
Postmortems can trace how failure propagated through dependencies and which bridge or hub amplified the cascade.
100. Network analysis and monitoring are connected
Monitoring edge loads, node states and topology changes can reveal emerging bottlenecks before service fails.
101. AI systems are networks of dependencies
Models, retrieval stores, tools, APIs, memory, permissions and users form an operational network. Failures can propagate across these dependencies.
102. Neural networks are not the same as generic network analysis
Artificial neural networks are computational models with specialised weighted connections and learning rules. Graph-theoretic network analysis can inspect them, but the concepts should not be conflated.
103. Knowledge graphs are explicit networks
Entities are nodes and typed relationships are edges. Graph structure supports retrieval, reasoning and provenance.
104. Citation networks reveal intellectual structure
Papers cite papers. Communities can represent research fields; central papers can bridge topics. Citation count alone does not equal scientific quality.
105. AI can help learners explore networks
Useful prompts include: “Create a network where the highest-degree node is not the highest-betweenness node,” “Show a cascade after one hub fails,” “Give me two community-detection partitions,” and “Challenge my choice of centrality metric.”
106. AI can invent edges
A generated network can look plausible while containing unsupported relationships. Every real scientific edge needs a defined evidence source.
107. Network extraction inherits text bias
If edges are extracted from publications, well-studied entities may appear more connected simply because they are mentioned more often.
108. Missing edges change centrality
Incomplete observation can make a node appear peripheral when its connections were simply not measured.
109. Sampling networks is difficult
Observing one person’s contacts can reveal new people outside the original sample. The sampling frame expands through the network.
110. Snowball sampling can overrepresent highly connected nodes
Because connected people are easier to reach through existing participants, naive estimates of degree and prevalence can be biased.
111. Respondent-driven approaches attempt corrections
They use assumptions about recruitment and degree to estimate population properties. Those assumptions require scrutiny.
112. Network null models matter
Is a clustering coefficient surprising relative to a random graph with the same number of nodes? Or relative to one preserving degree sequence? The reference network changes interpretation.
113. Erdős–Rényi-like random graphs provide one baseline
Edges occur independently with a common probability. Real networks often differ through degree heterogeneity, clustering or community structure.
114. Configuration models preserve degree sequence
They provide a stronger null when asking whether patterns exceed what degree distribution alone would produce.
115. Small-world structure combines local clustering with short paths
Some networks are highly clustered yet still globally reachable through a few long-range links.
116. Small-world claims require comparison
High clustering and short path length should be evaluated relative to suitable reference networks.
117. Percolation asks when connectivity collapses or emerges
As nodes or edges are added or removed, a giant connected component can appear or disappear around a critical region.
118. Percolation links network analysis to thresholds
Connectivity transitions are boundary conditions in network structure.
119. Dynamic rewiring can change resilience
Networks can adapt after disruption by creating new routes. Static analysis may underestimate recovery capability.
120. Adaptive networks couple state and topology
Node behaviour changes connections, and connections change node behaviour. Epidemics, social systems and infrastructure can show this feedback.
121. Network measures can be correlated
High degree and eigenvector centrality may often align. Reporting many centralities without a scientific reason can create redundant complexity.
122. Centrality rankings have uncertainty
When edges are uncertain or sampled incompletely, node rankings may change. Bootstrap or perturbation analysis can test stability.
123. Community assignments have uncertainty
Different algorithms or small edge changes may move nodes between communities. Stable conclusions should not depend on one arbitrary partition.
124. Network visualisations can mislead
Force-directed layouts place connected nodes near one another, but geometric distance on the picture may have no scientific meaning.
125. Node size and colour create rhetorical emphasis
A large red node looks important even before analysis. Visual encoding should reflect defined metrics transparently.
126. Edge crossing is often a layout artefact
Two drawn lines crossing does not mean the underlying network has an interaction at that crossing.
127. Parents can teach network thinking through everyday systems
Draw a family transport network, device network or study-concept network. Ask which connection is redundant, which is a bottleneck and what happens if one node fails.
128. Small-group tuition can use network diagnosis
Give students a concept map and remove one linking idea. Ask which questions become harder. This reveals prerequisite structure.
129. Examination performance benefits from networked knowledge
Transfer questions require connecting concepts across chapters. A learner with isolated topic islands struggles when unfamiliar questions demand cross-topic pathways.
130. The independence test
Give a new network without labels. Can the learner define nodes and edges, choose a meaningful centrality, identify communities and predict a cascade? That is transferable network reasoning.
131. The evidence boundary
Network structure can reveal possible pathways and dependence, but edges do not prove mechanism or causality automatically. Conclusions should match how edges were measured and what the network representation omits.
132. Independent-attempt task 1: centrality conflict
Draw two dense clusters joined by one bridge node. Identify highest degree and highest betweenness. Explain why the answers differ.
133. Independent-attempt task 2: community sensitivity
Remove two cross-community edges and predict how the partition changes.
134. Independent-attempt task 3: cascade
Create a capacity-limited network, remove one high-load edge and trace where overload moves next.
135. Independent-attempt task 4: null model
Ask whether ten triangles are surprising. Define two different random reference networks and explain why the answer may change.
136. Diagnostic error: degree equals importance
Repair by asking what kind of importance the scientific question needs.
137. Diagnostic error: edge meaning undefined
The graph shows lines but nobody knows whether they mean contact, correlation or causation. Repair by writing the edge contract.
138. Diagnostic error: missing direction
A flow network is analysed as undirected. Repair by restoring source-target meaning.
139. Diagnostic error: missing weights
One weak contact and one dominant pathway count equally. Repair by preserving edge strength where scientifically relevant.
140. Diagnostic error: cluster equals cause
Community structure is treated as proof of functional mechanism. Repair by testing external evidence.
141. Diagnostic error: one layout interpreted literally
Visual distance is mistaken for scientific distance. Repair by referring to graph path or measured spatial distance explicitly.
142. Diagnostic error: static network used for dynamic event
Connections changed during the process. Repair with temporal network data.
143. Diagnostic error: centrality without uncertainty
Rankings are treated as exact despite missing edges. Repair with sensitivity analysis.
144. A compact network-analysis checklist
- What are the nodes?
- What scientifically defines an edge?
- Is the network directed?
- Are edges weighted or signed?
- Is the network static or temporal?
- What sampling process generated the network?
- What edges or nodes may be missing?
- What kind of centrality matches the question?
- What paths or bottlenecks matter?
- Are there communities or modules?
- What null network is appropriate?
- How could diffusion or failure propagate?
- How robust is connectivity to perturbation?
- Are network metrics stable to plausible edge uncertainty?
- Does the network show association, mechanism or causality?
145. Frequently asked questions
What is scientific network analysis?
It is the quantitative study of entities and the relationships connecting them, focusing on structure such as paths, centrality, communities, diffusion and resilience.
What is centrality?
Centrality is a family of metrics describing different forms of structural importance in a network.
What is betweenness centrality?
It measures how often a node lies on shortest paths between other nodes, making it useful for identifying bridges or bottlenecks.
What is community detection?
It is the process of finding groups of nodes with denser or otherwise stronger internal connection structure than expected under a chosen model.
Does a network edge mean causation?
No. An edge means whatever relationship the data and network definition specify. Causal interpretation requires additional evidence.
How does network thinking help PSLE Science?
It strengthens food-web, circuit, system and interdependence reasoning by making connections and pathways explicit.
How does it deepen in Secondary Science?
Students can add centrality, weighted graphs, communities, diffusion, resilience, adjacency matrices and dynamic networks.
146. Continue the Science Education Systems series
- How Scientific Systems Thinking Works
- How Scientific Spatial Analysis Works
- How Scientific Robustness Works
Conclusion: In a network, the pattern of connection becomes part of the phenomenon
Maya sees the nodes.
Jia Jun defines the edges.
Hana asks which kind of centrality matters.
Ethan removes one bridge and watches the system reorganise.
Science needs all four.
Define the relationship.
preserve direction and weight.
measure paths.
find communities.
test cascades.
perturb the network.
Then let structure become evidence without allowing a beautiful graph to become a causal story by itself.
