Download as pdf or txt
Download as pdf or txt
You are on page 1of 26

Scale-free networks are rare

Jagoda Bobińska, Pasquale Gravante


What is a scale free network? power-law model
Why are they important?
Frequent claim:

Most real-world networks are scale-free!


Why are they important?
Frequent claim:

Most real-world networks are scale-free!

realistic simulations
Why are they important?
Frequent claim:

Most real-world networks are scale-free!

realistic simulations

studies - dynamics of processes in such a network


What if they are not?
Frequent claim:

Most real-world networks are scale-free!

1. Careful reassessment literature

2. Maybe it differs across domains?

3. How do real-world structures differ?


Why the ambiguity?
1. Small, domain-specific data sets

2. Insufficient statistical methods

3. Different definitions of “scale free”

4. Lack of comparison with other models


Research
ICON (Index of Complex Networks)

~1000 networks

biological transportation

500 informational technological 48

15 social 200
145
A network - procedure
A graph in a network - procedure
1. Graphʼs degree sequence:

2, 2, 2, 2, 3, 4, 4

2. Best-fitting power-law model for the sequence

3. Estimation how likely the power-law model is

4. Maybe it fits other models better?


5. Estimation how likely those models are
A graph in a network - procedure
1. Graphʼs degree sequence: R = LPL - LALT
2, 2, 2, 2, 3, 4, 4
1. R >> 0 - power-law model favoured
2. Best-fitting power-law model for the sequence
2. R << 0 - alternative model favoured

3. R ~ 0 - inconclusive
3. Estimation how likely the power-law model is (LPL)

4. Maybe it fits other models better?


5. Estimation how likely those models are (LALT)
Categorization
Super-weak

Weakest

Weakest
Weak

Weak
Strong

Strong
Strongest

Strongest
Super-weak
Method validation
RESULTS
HOW TO EVALUATE?

• Investigating the empirical evidence for universality of scale-free networks by


analyzing the estimated power-law scaling parameters Pr(α̂) across different data set
categories.

• Assessing the fit of the power-law distributions compared to alternatives.

• Combining metrics to provide a quantitative assessment of scale-free structure


evidence across various data sets.
A. POWER-LAW DISTRIBUTIONS

• A simple summary of each network data


set is the median scaling parameter α̂
among its corresponding simple graphs.

• Across the corpus, the overall distribution


of median parameters is concentrated
around a value of α̂ = 2, but also exhibits a
long right-tail.
A. POWER-LAW DISTRIBUTIONS

• 32% of data sets have α̂ ≥ 3, indicating a


substantial portion of networks with
heavy-tailed degree distributions.

• The range 2 < α̂ < 4 is often associated


with scale-free networks, with 43% of
data sets falling within this range.

• However, nearly 31% of networks have α̂


< 2, which is relatively unusual in the
scale-free network literature.
B. ALTERNATIVE DISTRIBUTIONS

• Independent of whether the power-law model is a statistically plausible explanation of a


particular degree sequence, comparing it with alternative distributions can be
informative.

• The likelihood ratio test can reveal whether some other distribution is an equally good or
even a better fit to the data.

• In our analysis of network data, we found only slight evidence favoring the power-law
distribution over alternative distributions based on likelihood ratio tests.
B. ALTERNATIVE DISTRIBUTIONS
• Exponential distribution: Favored over
the power law nearly as often as the
power law is favored over it.

• Log-normal distribution: Favored over


the power law more often than the
power law is favored over it, with tests
inconclusive in the remainder of cases. Power-law distribution with an exponential
cutoff: Just over half of the data sets favor the
• Weibull distribution: More often a better cutoff model over the pure power-law model,
statistical model of degree distributions and half of the times the test between the
than the power law. two is inconclusive.
C. ASSESSING THE SCALE-FREE HYPOTHESIS
• Each network dataset is classified according
to the evidence for scale-free structure that
was collected in the previous steps.

• 5 categories are created, ranging from


super-weak to strongest (evidence) plus an
additional one - Not Scale-Free at all.

• Results reveal the rarity of genuinely


scale-free networks, as they represent just the
4% of the datasets.
C. ASSESSING THE SCALE-FREE HYPOTHESIS

• 43% of network data sets classified as Not


Scale Free.

• Majority (52%) is categorized as Super-Weak.

• Only 11% and 4% classified as Strong and


Strongest, respectively, indicating very rare
occurrence of strong evidence for the
scale-free structure.

• Variability in scale-free evidence varies across


different network domains.
C. ASSESSING THE SCALE-FREE HYPOTHESIS
EVIDENCE BY DOMAIN
Biological Networks:
• 61% no evidence of scale-free structure, 35% indirect evidence.
• 6% strongest direct evidence, 8% strong direct evidence
Social Networks:
• 19% no evidence of scale-free structure 71% indirect evidence.
• No dataset in the Strong or Strongest categories.
Technological Networks:
• Only 7% lack evidence of scale-free structure.
• 92% exhibit indirect evidence.
• 28% exhibit strong direct evidence.
C. ASSESSING THE SCALE-FREE HYPOTHESIS

Example of a metabolic network:

This kind of network presents the strongest evidence of being


scale-free across all the dataset from the biological domain.

Example of a CAIDA network:

CAIDA datasets represent the connections between autonomous


systems on the Internet, they have the strongest evidence of being
scale-free across all the dataset from the technological domain.
CONCLUSIONS
• Scale-Free Networks Are Rare: The analysis of nearly 1000 real-world network datasets
revealed that genuinely scale-free networks are remarkably rare, with just around 4%
exhibiting the strongest evidence for scale-free structure.

• Lack of Universality: Contrary to common claims, scale-free structure is not universal


across different network domains. Instead, it varies considerably, with biological and
technological networks showing stronger evidence compared to social networks.

• Empirical Rarity: The empirical rarity of scale-free networks presents an opportunity for
future research. The focus on scale-free patterns has led to less exploration of
alternative mechanisms for generating realistic degree structures in networks.
IMPLICATIONS AND FUTURE DIRECTIONS

• Need for New Mechanisms: The scarcity of scale-free networks highlights the need for the
development and validation of new mechanisms that can generate more realistic degree
structures in networks, so, research should focus on exploring alternative models beyond
the traditional scale-free paradigm.

• Reassessment of Dynamical Processes: The diversity of degree structures in real-world


networks suggests that theoretical results concerning dynamical processes on networks,
may need to be reassessed to account for the structural diversity observed.

• Data-Driven Investigations: Future studies should use large datasets to evaluate the
actual claims in network science and gain new insights into network structure and function.
Discussion - method validation
Two methods of creating scale-free networks

One method for random graphs (Erdos-Renyi)


Thanks for your attention!

You might also like