threshold network models
DESCRIPTION
Presentation slides for a series of my academic done around 2005. The slides are rather mathematical.TRANSCRIPT
![Page 1: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/1.jpg)
Threshold network models
Naoki Masuda (Univ of Tokyo, Japan)
Refs:•Y. Ide, N. Konno, N. Masuda. Methodology & Computing in Applied Probability, 12, 361-377 (2010).•N. Masuda, N. Konno. Social Networks, 28, 297-309 (2006).•N. Konno, N. Masuda, R. Roy, A. Sarkar. J. Phys. A, 38, 6277-6291 (2005).•N. Masuda, H. Miwa, N. Konno. Phys. Rev. E, 71, 036108 (2005).•N. Masuda, H. Miwa, N. Konno. Phys. Rev. E, 70, 036124 (2004).
![Page 2: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/2.jpg)
Barabási & Albert model (1999)
• growing network• preferential attachment• power-law degree distribution. Called scale-free networks.
![Page 3: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/3.jpg)
Non-growing scale-free networks with intrinsic vertex weights
• Weight of node vi = wi. – wi is (i.i.d. and) distributed according to f(w).– Represents the propensity that vi gets edges.– Large wi ↔ large degree ki
• Generated nets become scale-free in many cases– Goh et al. (2001), Chung & Liu (2002), Caldarelli et al. (2002),
Söderberg (2002), Boguñá & Pastor-Satorras (2003) , etc.
• We investigate the threshold model (in Japanese, 閾値モデル ), which is one of such models, and its extensions.
vi and vj are connected ↔ wi + wj ≥ θ
![Page 4: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/4.jpg)
“meanfield” results (Caldarelli et al., 2002; Boguñá
& Pastor-Satorras, 2003)
• With f(w) a given weight distribution,
Degree distribution
1:1 relationship between k and w
(n: # vertices)
Cumulative dist. fn. of weight
degree
![Page 5: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/5.jpg)
Exponentially distributed weights (Caldarelli et al., 2002; Boguñá & Pastor-Sattoras, 2003)
Ave. deg. of neighbors
Vertex-wise clustering coef.
Degree dist.
Weight dist.
θ: threshold, n: # vertices
But real data often have C(k) ∝ k 1 (Vázquez et al., 2002; Ravasz et al., 2002, 2003)
: negative degree corr
✓ Good agreements with numerical results. ✓ Constraint w 0 is nonessential (cf. logistic dist)≧ ✓ Similar numerical results for Gaussian f(w)
degree
![Page 6: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/6.jpg)
Pareto distribution
• weight = city size, wealth distribution, etc.
✓ Good agreements with numerical results. ✓ Constraint w 0 is nonessential (cf. Cauchy dist)≧
![Page 7: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/7.jpg)
Mathematical definition as a random graph
(degree)
![Page 8: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/8.jpg)
Limit theorems for the degreeTheorem
(by SLLN for i.i.d. sequences)
Weak convergence corresponding to (2) can be shown by showing that the characteristic function of the LHS converges pointwiseto that of the RHS.
![Page 9: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/9.jpg)
Degree correlation
Proof: Calculate to see whether the characteristic function of the joint distribution of Dn(1)/n and Dn(2)/n {does/does not} factorize.
Theorem
• Dn(1)/n and Dn(2)/n are asymptotically independent.
• Given that vertices 1 and 2 are connected, Dn(1)/n and Dn(2)/n are not asymptotically independent.
![Page 10: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/10.jpg)
Limit theorems for # triangles
Theorem
✓ Extension to the case of larger “patterns” is straightforward.
standard normal var
a.s.
a.s.
![Page 11: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/11.jpg)
U‐statistics
: integrable, symmetric
![Page 12: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/12.jpg)
• g(r): prob two nodes with distance r are connected. – Internet: g(r) ≈ r 1 (Yook et al. PNAS 2002)
• We extend the threshold model.– Scatter nodes on (say) Rd
– Connect vi and vj iff (wi + wj) h(r) ≥ θ
– h(r) is nonincreasing. – Generally, g(r) ≠ h(r)
Spatial threshold model
![Page 13: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/13.jpg)
Drawing
![Page 14: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/14.jpg)
Spatial threshold model (i.e., (wi + wj) h(r) ≥ θ) generalizes
• (nonspatial) threshold model ← h(r) = 1
• Unit disk model ← wi = const
– Then, g(r) = 1[r ≤ rc]
• “Boolean model” (Meester & Roy, 1996) ← h(r) ≈ r 1
![Page 15: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/15.jpg)
In addition,• Gravity model (Zipf, 1947)
used in – Sociology (originally β=1)– Economics– Marketing
![Page 16: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/16.jpg)
Flavor of “physics” analysis
Example:
(wi + wj) h(r) ≥ θ(wi + wj) h(r) ≥ θ
1:1 relationship between k and w
degree
![Page 17: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/17.jpg)
Summary of the results
f(w) h(r) p(k) g(r) L
finite support
* finite support
finite support
large
λeλw r β stretched expon.
stretched expon.
large (if β is large)
λeλw (log r)1 k 1aβ/d r aβ small
∝ w –a–1 r β k 1aβ/d r aβ small (if aβ is small)
(wi + wj) h(r) ≥ θ(wi + wj) h(r) ≥ θ
✓ Good agreements with numerical results.
![Page 18: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/18.jpg)
Numerical results• (wi+wj) / r β ≥ θ• L is small for sufficiently small β.
– Phase transition at some βc?• C is large (i.e., many triangles)
– Show it analytically?
N=2000,4000,…,10000
Average path length (L) Clustering coefficient (C)
![Page 19: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/19.jpg)
Mathematics of the spatial threshold model
• Consider a homogeneous Poisson point process of intensity λ on Rd.
• Connect vi and vj iff
– Note: We consider only this h(r).
: enumeration of the point process
X0 X1
![Page 20: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/20.jpg)
• Degree of origin in a sphere of radius r :
• Cr(x) may converge or diverge as r → ∞
– Depending on f(w), θ, and β.– We consider (a representative of) each case.
Intuitively, = (degree of origin) / (volume of unit sphere)
Prob that a vertex with distance r from the origin is connected to the origin.
![Page 21: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/21.jpg)
Case 1: finite degree
where Δ is given via the characteristic function by
Volume of (d-1) dim unit sphere.
(convergence in distribution)
Theorem
![Page 22: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/22.jpg)
Sketch of proof
• Given X0 = x, Δr (i.e., degree of origin up to radius r) is Poisson with parameter
where
Prob that a vertex with distance r from the origin is connected to the origin.
& the dominated convergence theorem
~Volume of (d-1) dim unit sphere.
![Page 23: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/23.jpg)
Case 2: infinite degree
g(x): some function. Z: standard normal var
Example: β=1, d=2, (0 < α < 2, C > 0)
Sketch of proof: show that characteristic fn. of LHS converges to the product of two characteristic fns.
✓ direct calculations ✓ dominated convergence theorem
Theorem
![Page 24: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/24.jpg)
Thresholding + homophily• Hohophily: similar nodes (according to age, sex, education, race, etc.)
tends to be adjacent.• vi and vj are connected ↔
– wi + wj ≥ θ, and |wi - wj| ≤ c (or |wi - wj| / (wi + wj) ≤ c : Weber-Fechner law)
thresholding + homophilythresholding only
![Page 25: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/25.jpg)
ResultsThresholding + homophily
Homophily only
Thresholding only
w
k
k2
k2
k2
k
k
Degree dist
×: thresh + homo
■: thresh only
○: homo only
No longer hubs!
But still in a ‘special’ position
too many hubs
elites = hubs
f(w)= λexp(-λw)
![Page 26: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/26.jpg)
Some open problems
• (statistics of) number of triangles for the spatial threshold model
• Transition in average path length in the spatial threshold model
• General “thresholding function”– We have some results.
(wi + wj) h(r) ≥ θ(wi + wj) h(r) ≥ θ
e.g. h(r) = r –β
vi vj
![Page 27: Threshold network models](https://reader033.vdocuments.pub/reader033/viewer/2022051412/54c4b8004a7959f9778b464b/html5/thumbnails/27.jpg)
Conclusions• Threshold network model
– not growing– can be scale-free in many cases
• Extensions– spatial version– homophily and other interaction kernels
• References– Nonspatial threshold model
• Masuda, Miwa & Konno. Physical Review E, 70, 036124 (2004).
– Spatial threshold model• Masuda, Miwa & Konno. Physical Review E, 71, 036108 (2005).
– Limit theorems• Konno, Masuda, Roy & Sarkar. Journal of Physics A, 38, 6277 (2005)• Ide, Konno & Masuda, Methodology & Computing in Applied Probability, 12,
361 (2010).
– Homophily• Masuda & Konno. Social networks, 28, 297 (2006).