Deciphering and modeling heterogeneity in interaction networks

Transcription

1 Deciphering and modeling heterogeneity in interaction networks (using variational approximations) S. Robin INRA / AgroParisTech Mathematical Modeling of Complex Systems December 2013, Ecole Centrale de Paris

2 Outline State-space models for networks (inc. SBM) Variational inference (inc. SBM) (Variational) Bayesian model averaging Towards W -graphs

3 Heterogeneity in interaction networks

4 Understanding network structure Networks describe interactions between entities. Observed networks display heterogeneous topologies, that one would like to decipher and better understand.

5 Understanding network structure Networks describe interactions between entities. Observed networks display heterogeneous topologies, that one would like to decipher and better understand. Dolphine social network. Hyperlink network. [Newman and Girvan (2004)]

6 Modelling network heterogeneity Latent variable models allow to capture the underlying structure of a network. General setting for binary graphs. [Bollobás et al. (2007)]: A latent (unobserved) variable Z i is associated with each node: {Z i } iid π Edges Y ij = I{i j} are independent conditionally to the Z i s: {Y ij } independent {Z i } : Pr{Y ij = 1} = γ(z i, Z j ) We focus here on model approaches, in contrast with, e.g. Graph clustering [Girvan and Newman (2002)], [Newman (2004)]; Spectral clustering [von Luxburg et al. (2008)].

7 Modelling network heterogeneity Latent variable models allow to capture the underlying structure of a network. General setting for binary graphs. [Bollobás et al. (2007)]: A latent (unobserved) variable Z i is associated with each node: {Z i } iid π Edges Y ij = I{i j} are independent conditionally to the Z i s: {Y ij } independent {Z i } : Pr{Y ij = 1} = γ(z i, Z j ) We focus here on model approaches, in contrast with, e.g. Graph clustering [Girvan and Newman (2002)], [Newman (2004)]; Spectral clustering [von Luxburg et al. (2008)].

8 Modelling network heterogeneity Latent variable models allow to capture the underlying structure of a network. General setting for binary graphs. [Bollobás et al. (2007)]: A latent (unobserved) variable Z i is associated with each node: {Z i } iid π Edges Y ij = I{i j} are independent conditionally to the Z i s: {Y ij } independent {Z i } : Pr{Y ij = 1} = γ(z i, Z j ) We focus here on model approaches, in contrast with, e.g. Graph clustering [Girvan and Newman (2002)], [Newman (2004)]; Spectral clustering [von Luxburg et al. (2008)].

9 Latent space models State-space model: principle.

10 Latent space models State-space model: principle. Consider n nodes (i = 1..n);

11 Latent space models State-space model: principle. Consider n nodes (i = 1..n); Z i = unobserved position of node i, e.g. {Z i } iid N (0, I )

12 Latent space models State-space model: principle. Consider n nodes (i = 1..n); Z i = unobserved position of node i, e.g. {Z i } iid N (0, I ) Edge {Y ij } independent given {Z i }, e.g. Pr{Y ij = 1} = γ(z i, Z j ) γ(z, z ) = f ( z, z ).

13 Latent space models State-space model: principle. Consider n nodes (i = 1..n); Z i = unobserved position of node i, e.g. {Z i } iid N (0, I ) Y = Edge {Y ij } independent given {Z i }, e.g. Pr{Y ij = 1} = γ(z i, Z j ) γ(z, z ) = f ( z, z ).

14 A variety of state-space models Continuous. Latent position models. [Hoff et al. (2002)]: Z i R d, logit[γ(z, z )] = a z z [Handcock et al. (2007)]: Z i k p k N d (µ k, σ 2 k I ) [Lovász and Szegedy (2006)]: Z i U [0,1], γ(z, z ) : [0, 1] 2 [0, 1] =: graphon function [Daudin et al. (2010)]: Z i S K, γ(z, z ) = k,l z k z l γ kl

15 Stochastic Block Model (SBM) A mixture model for random graphs. [Nowicki and Snijders (2001)]

16 Stochastic Block Model (SBM) A mixture model for random graphs. [Nowicki and Snijders (2001)] Consider n nodes (i = 1..n);

17 Stochastic Block Model (SBM) A mixture model for random graphs. [Nowicki and Snijders (2001)] Consider n nodes (i = 1..n); Z i = unobserved label of node i: {Z i } iid M(1; π) π = (π 1,...π K );

18 Stochastic Block Model (SBM) A mixture model for random graphs. [Nowicki and Snijders (2001)] Consider n nodes (i = 1..n); Z i = unobserved label of node i: {Z i } iid M(1; π) π = (π 1,...π K ); Edge Y ij depends on the labels: {Y ij } independent given {Z i }, Pr{Y ij = 1} = γ(z i, Z j )

19 Stochastic Block Model (SBM) A mixture model for random graphs. [Nowicki and Snijders (2001)] Consider n nodes (i = 1..n); Z i = unobserved label of node i: {Z i } iid M(1; π) π = (π 1,...π K ); Edge Y ij depends on the labels: {Y ij } independent given {Z i }, Pr{Y ij = 1} = γ(z i, Z j )

20 Variational inference Joint work with J.-J. Daudin, S. Gazal, F. Picard [Daudin et al. (2008)], [Gazal et al. (2012)]

21 Incomplete data models Aim. Based on the observed network Y = (Y ij ), we want to infer the parameters θ = (π, γ) the hidden states Z = (Z i ) State space models belong to the class of incomplete data models as the edges (Y ij ) are observed, the latent positions (or status) (Z i ) are not. usual issue in unsupervised classification.

22 Incomplete data models Aim. Based on the observed network Y = (Y ij ), we want to infer the parameters θ = (π, γ) the hidden states Z = (Z i ) State space models belong to the class of incomplete data models as the edges (Y ij ) are observed, the latent positions (or status) (Z i ) are not. usual issue in unsupervised classification.

23 Frequentist maximum likelihood inference Likelihood. The (log-)likelihood log P(Y ; θ) = log Z P(Y, Z; θ) can not be computed. EM trick. But it can be decomposed as log P(Y ; θ) = log P(Y, Z; θ) log P(Z Y ; θ), the conditional expectation of which gives E[log P(Y ; θ) Y ] = E[log P(Y, Z; θ) Y ] E[log P(Z Y ; θ) Y ] log P(Y ; θ) = E[log P(Y, Z; θ) Y ] + H[P(Z Y ; θ)] where H stands for the entropy.

24 Frequentist maximum likelihood inference Likelihood. The (log-)likelihood log P(Y ; θ) = log Z P(Y, Z; θ) can not be computed. EM trick. But it can be decomposed as log P(Y ; θ) = log P(Y, Z; θ) log P(Z Y ; θ), the conditional expectation of which gives E[log P(Y ; θ) Y ] = E[log P(Y, Z; θ) Y ] E[log P(Z Y ; θ) Y ] log P(Y ; θ) = E[log P(Y, Z; θ) Y ] + H[P(Z Y ; θ)] where H stands for the entropy.

25 Frequentist maximum likelihood inference Likelihood. The (log-)likelihood log P(Y ; θ) = log Z P(Y, Z; θ) can not be computed. EM trick. But it can be decomposed as log P(Y ; θ) = log P(Y, Z; θ) log P(Z Y ; θ), the conditional expectation of which gives E[log P(Y ; θ) Y ] = E[log P(Y, Z; θ) Y ] E[log P(Z Y ; θ) Y ] log P(Y ; θ) = E[log P(Y, Z; θ) Y ] + H[P(Z Y ; θ)] where H stands for the entropy.

26 Frequentist maximum likelihood inference Likelihood. The (log-)likelihood log P(Y ; θ) = log Z P(Y, Z; θ) can not be computed. EM trick. But it can be decomposed as log P(Y ; θ) = log P(Y, Z; θ) log P(Z Y ; θ), the conditional expectation of which gives E[log P(Y ; θ) Y ] = E[log P(Y, Z; θ) Y ] E[log P(Z Y ; θ) Y ] log P(Y ; θ) = E[log P(Y, Z; θ) Y ] + H[P(Z Y ; θ)] where H stands for the entropy.

27 EM algorithm Aims at maximizing the log-likelihood log P(Y ; θ) through the alternation of two steps [Dempster et al. (1977)] M-step: maximize E[log P(Y, Z; θ) Y ] with respect to θ generally similar to standard MLE. E-step: calculate P(Z Y ) (at least, some moments) sometimes straightforward (independent mixture models: Bayes formula) sometimes tricky but doable (HMMs: forward-backward recursion) sometimes impossible (SBM:...)

28 EM algorithm Aims at maximizing the log-likelihood log P(Y ; θ) through the alternation of two steps [Dempster et al. (1977)] M-step: maximize E[log P(Y, Z; θ) Y ] with respect to θ generally similar to standard MLE. E-step: calculate P(Z Y ) (at least, some moments) sometimes straightforward (independent mixture models: Bayes formula) sometimes tricky but doable (HMMs: forward-backward recursion) sometimes impossible (SBM:...)

29 EM algorithm Aims at maximizing the log-likelihood log P(Y ; θ) through the alternation of two steps [Dempster et al. (1977)] M-step: maximize E[log P(Y, Z; θ) Y ] with respect to θ generally similar to standard MLE. E-step: calculate P(Z Y ) (at least, some moments) sometimes straightforward (independent mixture models: Bayes formula) sometimes tricky but doable (HMMs: forward-backward recursion) sometimes impossible (SBM:...)

30 EM algorithm Aims at maximizing the log-likelihood log P(Y ; θ) through the alternation of two steps [Dempster et al. (1977)] M-step: maximize E[log P(Y, Z; θ) Y ] with respect to θ generally similar to standard MLE. E-step: calculate P(Z Y ) (at least, some moments) sometimes straightforward (independent mixture models: Bayes formula) sometimes tricky but doable (HMMs: forward-backward recursion) sometimes impossible (SBM:...)

31 EM algorithm Aims at maximizing the log-likelihood log P(Y ; θ) through the alternation of two steps [Dempster et al. (1977)] M-step: maximize E[log P(Y, Z; θ) Y ] with respect to θ generally similar to standard MLE. E-step: calculate P(Z Y ) (at least, some moments) sometimes straightforward (independent mixture models: Bayes formula) sometimes tricky but doable (HMMs: forward-backward recursion) sometimes impossible (SBM:...)

32 EM algorithm Aims at maximizing the log-likelihood log P(Y ; θ) through the alternation of two steps [Dempster et al. (1977)] M-step: maximize E[log P(Y, Z; θ) Y ] with respect to θ generally similar to standard MLE. E-step: calculate P(Z Y ) (at least, some moments) sometimes straightforward (independent mixture models: Bayes formula) sometimes tricky but doable (HMMs: forward-backward recursion) sometimes impossible (SBM:...)

33 Conditional distribution of Z EM works provided that we can calculate P(Z Y ; θ)

34 Conditional distribution of Z EM works provided that we can calculate P(Z Y ; θ)

35 Conditional distribution of Z EM works provided that we can calculate P(Z Y ; θ)

36 Conditional distribution of Z EM works provided that we can calculate P(Z Y ; θ)

37 Conditional distribution of Z EM works provided that we can calculate P(Z Y ; θ) but we can not for state space models for graph.

38 Conditional distribution of Z EM works provided that we can calculate P(Z Y ; θ) but we can not for state space models for graph. Conditional distribution. The dependency graph of Z given Y is a clique. No factorization can be hoped (unlike for HMM). P(Z Y ; θ) can not be computed (efficiently). Variational techniques may help as they provide Q(Z) P(Z Y ).

39 Variational inference Lower bound of the log-likelihood. For any distribution Q(Z) [Jaakkola (2000),Wainwright and Jordan (2008)], log P(Y ) log P(Y ) KL[Q(Z), P(Z Y )] = Q(Z) log P(Y, Z) dz Q(Z) log Q(Z) dz = E Q [log P(Y, Z)] + H[Q(Z)] Link with EM. This is similar to log P(Y ) = E[log P(Y, Z) Y ] + H[P(Z Y )] replacing P(Z Y ) with Q(Z).

40 Variational inference Lower bound of the log-likelihood. For any distribution Q(Z) [Jaakkola (2000),Wainwright and Jordan (2008)], log P(Y ) log P(Y ) KL[Q(Z), P(Z Y )] = Q(Z) log P(Y, Z) dz Q(Z) log Q(Z) dz = E Q [log P(Y, Z)] + H[Q(Z)] Link with EM. This is similar to log P(Y ) = E[log P(Y, Z) Y ] + H[P(Z Y )] replacing P(Z Y ) with Q(Z).

41 Variational inference Lower bound of the log-likelihood. For any distribution Q(Z) [Jaakkola (2000),Wainwright and Jordan (2008)], log P(Y ) log P(Y ) KL[Q(Z), P(Z Y )] = Q(Z) log P(Y, Z) dz Q(Z) log Q(Z) dz = E Q [log P(Y, Z)] + H[Q(Z)] Link with EM. This is similar to log P(Y ) = E[log P(Y, Z) Y ] + H[P(Z Y )] replacing P(Z Y ) with Q(Z).

42 Variational inference Lower bound of the log-likelihood. For any distribution Q(Z) [Jaakkola (2000),Wainwright and Jordan (2008)], log P(Y ) log P(Y ) KL[Q(Z), P(Z Y )] = Q(Z) log P(Y, Z) dz Q(Z) log Q(Z) dz = E Q [log P(Y, Z)] + H[Q(Z)] Link with EM. This is similar to log P(Y ) = E[log P(Y, Z) Y ] + H[P(Z Y )] replacing P(Z Y ) with Q(Z).

43 Variational EM algorithm Variational EM. M-step: compute θ = arg max E Q [log P(Y, Z; θ)]. θ E-step: replace the calculation of P(Z Y ) with the search of Q (Z) = arg min KL[Q(Z), P(Z Y )]. Q Q Taking Q = {all possible distributions} gives Q (Z) = P(Z Y )... like EM does. Variational approximations rely on the choice a set Q of good and tractable ditributions.

44 Variational EM algorithm Variational EM. M-step: compute θ = arg max E Q [log P(Y, Z; θ)]. θ E-step: replace the calculation of P(Z Y ) with the search of Q (Z) = arg min KL[Q(Z), P(Z Y )]. Q Q Taking Q = {all possible distributions} gives Q (Z) = P(Z Y )... like EM does. Variational approximations rely on the choice a set Q of good and tractable ditributions.

45 Variational EM algorithm Variational EM. M-step: compute θ = arg max E Q [log P(Y, Z; θ)]. θ E-step: replace the calculation of P(Z Y ) with the search of Q (Z) = arg min KL[Q(Z), P(Z Y )]. Q Q Taking Q = {all possible distributions} gives Q (Z) = P(Z Y )... like EM does. Variational approximations rely on the choice a set Q of good and tractable ditributions.

46 Variational EM for SBM Distribution class. Q = set of factorisable distributions: Q = {Q : Q(Z) = i Q i (Z i )}, Q i (Z i ) = k τ Z ik ik. The approximate joint distribution is Q(Z i, Z j ) = Q i (Z i )Q j (Z j ). The optimal approximation within this class satisfies a fix-point relation: τ ik π k f kl (Y ij ) τ jl j i l also known as mean-field approximation in physics [Parisi (1988)]. Variational estimates. No general statistical guaranty for variational estimates. SBM is a very specific case for which the variational approximation is asymptotically exact [Celisse et al. (2012),Mariadassou and Matias (2012)].

47 Variational EM for SBM Distribution class. Q = set of factorisable distributions: Q = {Q : Q(Z) = i Q i (Z i )}, Q i (Z i ) = k τ Z ik ik. The approximate joint distribution is Q(Z i, Z j ) = Q i (Z i )Q j (Z j ). The optimal approximation within this class satisfies a fix-point relation: τ ik π k f kl (Y ij ) τ jl j i l also known as mean-field approximation in physics [Parisi (1988)]. Variational estimates. No general statistical guaranty for variational estimates. SBM is a very specific case for which the variational approximation is asymptotically exact [Celisse et al. (2012),Mariadassou and Matias (2012)].

48 Variational EM for SBM Distribution class. Q = set of factorisable distributions: Q = {Q : Q(Z) = i Q i (Z i )}, Q i (Z i ) = k τ Z ik ik. The approximate joint distribution is Q(Z i, Z j ) = Q i (Z i )Q j (Z j ). The optimal approximation within this class satisfies a fix-point relation: τ ik π k f kl (Y ij ) τ jl j i l also known as mean-field approximation in physics [Parisi (1988)]. Variational estimates. No general statistical guaranty for variational estimates. SBM is a very specific case for which the variational approximation is asymptotically exact [Celisse et al. (2012),Mariadassou and Matias (2012)].

49 Variational Bayes inference Bayesian perspective. θ = (π, γ) is random. Model:

50 Variational Bayes inference Bayesian perspective. θ = (π, γ) is random. Model: P(θ)

51 Variational Bayes inference Bayesian perspective. θ = (π, γ) is random. Model: P(θ) P(Z π)

52 Variational Bayes inference Bayesian perspective. θ = (π, γ) is random. Model: P(θ) P(Z π) P(Y γ, Z)

53 Variational Bayes inference Bayesian perspective. θ = (π, γ) is random. Model: P(θ) P(Z π) P(Y γ, Z) Bayesian inference. The aim is then to get the joint conditional distribution of the parameters and of the hidden variables: P(θ, Z Y ).

54 Variational Bayes algorithm Variational Bayes. As P(θ, Z Y ) is intractable, one look formula Q (θ, Z) = arg min KL [Q(θ, Z) P(θ, Z Y )] Q Q taking, e.g., Q = {Q(θ, Z) = Q θ (θ)q Z (Z)} Variational Bayes EM (VBEM). When P(Z, Y θ) belongs to the exponential family, P(θ) is the corresponding conjugate prior, Q can be obtained iteratively as [Beal and Ghahramani (2003)] log Q h θ (θ) E Q h 1 Z [log P(Z, Y, θ)], log QZ h (Z) E Qθ h [log P(Z, Y, θ)]. Application to SBM: [Latouche et al. (2012)]

55 Variational Bayes algorithm Variational Bayes. As P(θ, Z Y ) is intractable, one look formula Q (θ, Z) = arg min KL [Q(θ, Z) P(θ, Z Y )] Q Q taking, e.g., Q = {Q(θ, Z) = Q θ (θ)q Z (Z)} Variational Bayes EM (VBEM). When P(Z, Y θ) belongs to the exponential family, P(θ) is the corresponding conjugate prior, Q can be obtained iteratively as [Beal and Ghahramani (2003)] log Q h θ (θ) E Q h 1 Z [log P(Z, Y, θ)], log QZ h (Z) E Qθ h [log P(Z, Y, θ)]. Application to SBM: [Latouche et al. (2012)]

56 Variational Bayes algorithm Variational Bayes. As P(θ, Z Y ) is intractable, one look formula Q (θ, Z) = arg min KL [Q(θ, Z) P(θ, Z Y )] Q Q taking, e.g., Q = {Q(θ, Z) = Q θ (θ)q Z (Z)} Variational Bayes EM (VBEM). When P(Z, Y θ) belongs to the exponential family, P(θ) is the corresponding conjugate prior, Q can be obtained iteratively as [Beal and Ghahramani (2003)] log Q h θ (θ) E Q h 1 Z [log P(Z, Y, θ)], log QZ h (Z) E Qθ h [log P(Z, Y, θ)]. Application to SBM: [Latouche et al. (2012)]

57 VBEM: Simulation study Credibility intervals: π 1 : +, γ 11 :, γ 12 :, γ 22 : Width of the posterior credibility intervals. π 1, γ 11, γ 12, γ 22

58 VBEM: Simulation study Credibility intervals: π 1 : +, γ 11 :, γ 12 :, γ 22 : Width of the posterior credibility intervals. π 1, γ 11, γ 12, γ 22

59 SBM analysis of E. coli operon networks [Picard et al. (2009)]

60 SBM analysis of E. coli operon networks Meta-graph representation. [Picard et al. (2009)]

61 SBM analysis of E. coli operon networks Meta-graph representation. Parameter estimates. K = 5 [Picard et al. (2009)]

62 (Variational) Bayesian model averaging Joint work with M.-L. Martin-Magniette, S. Volant [Volant et al. (2012)]

63 Model choice Model selection. The number of classes K generally needs to be estimated. In the frequentist setting, an approximate ICL criterion can be derived ICL = E[log P(Y, Z) Y ] 1 { K(K + 1) n(n 1) log (K 1) log n In the Bayesian setting, exact versions of BIC and ICL criteria can be calculated as log P(Y, K), log P(Y, Z, K). (up to the variational approximation) [Latouche et al. (2012)] But, in some applications, it may be useless or meaningless and model averaging may be preferred.

64 Model choice Model selection. The number of classes K generally needs to be estimated. In the frequentist setting, an approximate ICL criterion can be derived ICL = E[log P(Y, Z) Y ] 1 { K(K + 1) n(n 1) log (K 1) log n In the Bayesian setting, exact versions of BIC and ICL criteria can be calculated as log P(Y, K), log P(Y, Z, K). (up to the variational approximation) [Latouche et al. (2012)] But, in some applications, it may be useless or meaningless and model averaging may be preferred.

65 Bayesian model averaging (BMA) General principle. [Hoeting et al. (1999)] : a parameter that can be defined under a series of different models {M K } K. Denote P K ( Y ) its posterior distribution under model M K. The posterior distribution of can be averaged over all models as P( Y ) = K P(M K Y )P K ( Y ) Remarks. w k = P(M K Y ): weight given to model M K for the estimation of. Calculating of w K is not easy, but variational approximation may help.

66 Bayesian model averaging (BMA) General principle. [Hoeting et al. (1999)] : a parameter that can be defined under a series of different models {M K } K. Denote P K ( Y ) its posterior distribution under model M K. The posterior distribution of can be averaged over all models as P( Y ) = K P(M K Y )P K ( Y ) Remarks. w k = P(M K Y ): weight given to model M K for the estimation of. Calculating of w K is not easy, but variational approximation may help.

67 Variational Bayesian model averaging (VBMA) Variational Bayes formulation. M K can be viewed as one more hidden layer Variational Bayes then aims at finding Q (K, θ, Z) = arg min KL [Q(K, θ, Z) P(K, θ, Z Y )] Q Q with Q = {Q(θ, Z, K) = Q θ (θ K)Q Z (Z K)Q K (K)} 1 Optimal variational weights: Q K (K) P(K) exp{log P(Y K) KL[Q (Z, θ K); P(Z, θ Y, K)]} = P(K Y )e KL[Q (Z,θ K);P(Z,θ Y,K)]. 1 No additional approximation w.r.t. the regular VBEM.

68 Variational Bayesian model averaging (VBMA) Variational Bayes formulation. M K can be viewed as one more hidden layer Variational Bayes then aims at finding Q (K, θ, Z) = arg min KL [Q(K, θ, Z) P(K, θ, Z Y )] Q Q with Q = {Q(θ, Z, K) = Q θ (θ K)Q Z (Z K)Q K (K)} 1 Optimal variational weights: Q K (K) P(K) exp{log P(Y K) KL[Q (Z, θ K); P(Z, θ Y, K)]} = P(K Y )e KL[Q (Z,θ K);P(Z,θ Y,K)]. 1 No additional approximation w.r.t. the regular VBEM.

69 VBMA: the recipe For K = 1... K max Use regular VBEM to compute the approximate conditional posterior Q Z,θ K (Z, θ K) = Q Z K (Z K)Q θ K (θ K); Compute the conditional lower bound of the log-likelihood L K = log P(Y K) KL[Q (Z, θ K); P(Z, θ Y, K)]. Compute the approximate posterior of model M K : w K := Q K (K) P(K) exp(l K ). Deduce the approximate posterior of : P( Y ) K w K Q θ K ( K)

70 VBMA: the recipe For K = 1... K max Use regular VBEM to compute the approximate conditional posterior Q Z,θ K (Z, θ K) = Q Z K (Z K)Q θ K (θ K); Compute the conditional lower bound of the log-likelihood L K = log P(Y K) KL[Q (Z, θ K); P(Z, θ Y, K)]. Compute the approximate posterior of model M K : w K := Q K (K) P(K) exp(l K ). Deduce the approximate posterior of : P( Y ) K w K Q θ K ( K)

71 VBMA: the recipe For K = 1... K max Use regular VBEM to compute the approximate conditional posterior Q Z,θ K (Z, θ K) = Q Z K (Z K)Q θ K (θ K); Compute the conditional lower bound of the log-likelihood L K = log P(Y K) KL[Q (Z, θ K); P(Z, θ Y, K)]. Compute the approximate posterior of model M K : w K := Q K (K) P(K) exp(l K ). Deduce the approximate posterior of : P( Y ) K w K Q θ K ( K)

72 Towards W -graphs Joint work with P. Latouche. [Latouche and Robin (2013)]

73 W -graph model W graph. Latent variables: Graphon function γ(z, z ) (Z i ) iid U [0,1], Graphon function γ: γ(z, z ) : [0, 1] 2 [0, 1] Edges: Pr{Y ij = 1} = γ(z i, Z j )

74 Inference of the graphon function Probabilistic point of view. W -graph have been mostly studied in the probability literature: [Lovász and Szegedy (2006)], [Diaconis and Janson (2008)] Motif (sub-graph) frequencies are invariant characteristics of a W -graph. Intrinsic un-identifiability of the graphon function γ is often overcome by imposing that u γ(u, v) dv is monotonous increasing. Statistical point of view. Not much attention has been paid to its inference until very recently: [Chatterjee (2012)], [Airoldi et al. (2013)], [Wolfe and Olhede (2013)],... The two latter also uses SBM as a proxy for W -graph.

75 Inference of the graphon function Probabilistic point of view. W -graph have been mostly studied in the probability literature: [Lovász and Szegedy (2006)], [Diaconis and Janson (2008)] Motif (sub-graph) frequencies are invariant characteristics of a W -graph. Intrinsic un-identifiability of the graphon function γ is often overcome by imposing that u γ(u, v) dv is monotonous increasing. Statistical point of view. Not much attention has been paid to its inference until very recently: [Chatterjee (2012)], [Airoldi et al. (2013)], [Wolfe and Olhede (2013)],... The two latter also uses SBM as a proxy for W -graph.

76 SBM as a W -graph model Latent variables: (U i ) iid U[0, 1] Graphon function γ SBM K (z, z ) Z ik = I{σ k 1 U i < σ k } where σ k = k l=1 π l. Blockwise constant graphon: Edges: γ(z, z ) = γ kl Pr{Y ij = 1} = γ(z i, Z j )

77 Variational Bayes estimation of γ(z, z ) Posterior mean of γk SBM (z, z ) VBEM inference provides the approximate posteriors: (π Y ) Dir(π ) (γ kl Y ) Beta(γkl 0 kl ) Estimate of γ(u, v). γ SBM K (u, v) = Ẽ ( γ C(u),C(v) Y ) where C(u) = 1 + k I{σ k u}. [Gouda and Szántai (2010)]

78 Model averaging Model averaging: There is no true K in the W -graph model. Apply VBMA recipe. For K = 1..K max, fit an SBM model via VBEM and compute Perform model averaging as γ SBM K (z, z ) = E Q [γ C(z),C(z )]. γ(z, z ) = K w K γ SBM K (z, z ) where w K is the variational weights arising from variational Bayes inference.

79 Model averaging Model averaging: There is no true K in the W -graph model. Apply VBMA recipe. For K = 1..K max, fit an SBM model via VBEM and compute Perform model averaging as γ SBM K (z, z ) = E Q [γ C(z),C(z )]. γ(z, z ) = K w K γ SBM K (z, z ) where w K is the variational weights arising from variational Bayes inference.

80 Some simulations Design. Symetric graphon: γ(u, v) = ρλ 2 (uv) λ 1 Variational posterior for K: Q (K). λ : imbalanced graph ρ : dense graph Results. More complex models as n and λ Posterior fairly concentrated

81 French political blogosphere Website network. French political blogs: 196 nodes, 1432 edges.

82 French political blogosphere Infered graphon. Ŵ (u, v) = E(γ(u, v) Y ) Motif probability can be estimated as well as µ(m) = E(µ(m) Y ).

83 French political blogosphere Infered graphon. Ŵ (u, v) = E(γ(u, v) Y ) Motif probability can be estimated as well as µ(m) = E(µ(m) Y ).

84 Conclusion Real networks are heterogeneous. State space models (including SBM) allow to capture such an heterogeneity... but raise inference problems. Variational approximations help to deal with complex (conditional) dependency structures. SBM is a special (rare?) case for which guaranties exist as for the performances of the variational approximation. SBM can be used as a proxy for smoother models, such as W -graphs.

85 Conclusion Real networks are heterogeneous. State space models (including SBM) allow to capture such an heterogeneity... but raise inference problems. Variational approximations help to deal with complex (conditional) dependency structures. SBM is a special (rare?) case for which guaranties exist as for the performances of the variational approximation. SBM can be used as a proxy for smoother models, such as W -graphs.

86 Conclusion Real networks are heterogeneous. State space models (including SBM) allow to capture such an heterogeneity... but raise inference problems. Variational approximations help to deal with complex (conditional) dependency structures. SBM is a special (rare?) case for which guaranties exist as for the performances of the variational approximation. SBM can be used as a proxy for smoother models, such as W -graphs.

87 Conclusion Real networks are heterogeneous. State space models (including SBM) allow to capture such an heterogeneity... but raise inference problems. Variational approximations help to deal with complex (conditional) dependency structures. SBM is a special (rare?) case for which guaranties exist as for the performances of the variational approximation. SBM can be used as a proxy for smoother models, such as W -graphs.

88 Conclusion Real networks are heterogeneous. State space models (including SBM) allow to capture such an heterogeneity... but raise inference problems. Variational approximations help to deal with complex (conditional) dependency structures. SBM is a special (rare?) case for which guaranties exist as for the performances of the variational approximation. SBM can be used as a proxy for smoother models, such as W -graphs.

89 Airoldi, E. M., Costa, T. B. and Chan, S. H. (Nov., 2013). Stochastic blockmodel approximation of a graphon: Theory and consistent estimation. ArXiv e-prints Beal, J., M. and Ghahramani, Z. (2003). The variational Bayesian EM algorithm for incomplete data: with application to scoring graphical model structures. Bayes. Statist Bollobás, B., Janson, S. and Riordan, O. (2007). The phase transition in inhomogeneous random graphs. Rand. Struct. Algo. 31 (1) Celisse, A., Daudin, J.-J. and Pierre, L. (2012). Consistency of maximum-likelihood and variational estimators in the stochastic block model. Electron. J. Statis Chatterjee, S. (Dec., 2012). Matrix estimation by Universal Singular Value Thresholding. ArXiv e-prints Daudin, J.-J., Picard, F. and Robin, S. (Jun, 2008). A mixture model for random graphs. Stat. Comput. 18 (2) Daudin, J.-J., Pierre, L. and Vacher, C. (2010). Model for heterogeneous random networks using continuous latent variables and an application to a tree fungus network. Biometrics. 66 (4) Dempster, A. P., Laird, N. M. and Rubin, D. B. (1977). Maximum likelihood from incomplete data via the EM algorithm. J. R. Statist. Soc. B Diaconis, P. and Janson, S. (2008). Graph limits and exchangeable random graphs. Rend. Mat. Appl. 7 (28) Gazal, S., Daudin, J.-J. and Robin, S. (2012). Accuracy of variational estimates for random graph mixture models. Journal of Statistical Computation and Simulation. 82 (6) Girvan, M. and Newman, M. E. J. (2002). Community strucutre in social and biological networks. Proc. Natl. Acad. Sci. USA. 99 (12) Gouda, A. and Szántai, T. (2010). On numerical calculation of probabilities according to Dirichlet distribution. Ann. Oper. Res DOI: /s

90 Handcock, M., Raftery, A. and Tantrum, J. (2007). Model-based clustering for social networks. JRSSA. 170 (2) doi: /j X x. Hoeting, J. A., Madigan, D., Raftery, A. E. and Volinsky, C. T. (1999). Bayesian model averaging: A tutorial. Statistical Science. 14 (4) Hoff, P. D., Raftery, A. E. and Handcock, M. S. (2002). Latent space approaches to social network analysis. J. Amer. Statist. Assoc. 97 (460) Jaakkola, T. (2000). Advanced mean field methods: theory and practice. chapter Tutorial on variational approximation methods. MIT Press. Latouche, P., Birmelé, E. and Ambroise, C. (2012). Variational bayesian inference and complexity control for stochastic block models. Statis. Model. 12 (1) Latouche, P. and Robin, S. (Oct., 2013). Bayesian Model Averaging of Stochastic Block Models to Estimate the Graphon Function and Motif Frequencies in a W-graph Model. ArXiv e-prints Lovász, L. and Szegedy, B. (2006). Limits of dense graph sequences. Journal of Combinatorial Theory, Series B. 96 (6) von Luxburg, U., Belkin, M. and Bousquet, O. (2008). Consistency of spectral clustering. Ann. Stat. 36 (2) Mariadassou, M. and Matias, C. (June, 2012). Convergence of the groups posterior distribution in latent or stochastic block models. ArXiv e-prints Newman, M. and Girvan, M. (2004). Finding and evaluating community structure in networks,. Phys. Rev. E Newman, M. E. J. (2004). Fast algorithm for detecting community structure in networks. Phys. Rev. E (69) Nowicki, K. and Snijders, T. (2001). Estimation and prediction for stochastic block-structures. J. Amer. Statist. Assoc Parisi, G. (1988). Statistical Field Theory. Addison Wesley, New York),.

91 Picard, F., Miele, V., Daudin, J.-J., Cottret, L. and Robin, S. (2009). Deciphering the connectivity structure of biological networks using mixnet. BMC Bioinformatics. Suppl 6 S17. doi: / s6-s17. Volant, S., Magniette, M.-L. M. and Robin, S. (2012). Variational bayes approach for model aggregation in unsupervised classification with markovian dependency. Comput. Statis. & Data Analysis. 56 (8) Wainwright, M. J. and Jordan, M. I. (2008). Graphical models, exponential families, and variational inference. Found. Trends Mach. Learn. 1 (1 2) Wolfe, P. J. and Olhede, S. C. (Sep., 2013). Nonparametric graphon estimation. ArXiv e-prints

92 SBM for a binary social network Zachary data. Social binary network of friendship within a sport club. (%) γ kl k/l π l

93 Küllback-Leibler divergence Definition. KL[Q( ), P( )] = = Q(Z) log Q(Z) P(Z) dz Q(Z) log Q(Z) dz Q(Z) log P(Z) dz Some properties. Always positive. Null iff P = Q. Contrast consistent with maximum-likelihood inference. Not a distance, only a divergence.

94 Concentration of the degree distribution n = 100.

95 Concentration of the degree distribution n = 1000.

96 Concentration of the degree distribution n =