ChapterPDF Available

Learning of Weighted Multi-layer Networks via Dynamic Social Spaces, with Application to Financial Interbank Transactions


Abstract and Figures

We propose a general network model suited for longitudinal data of multi-layer networks with directed and weighted edges. Our formulation built upon the latent social space representation of networks. It consists of a hierarchical formulation: deep levels of the model represent latent coordinates of agents in the social space, evolving in continuous time via Gaussian Processes; meanwhile, top levels jointly manage incidence and strength of interactions by considering a Zero-Inflated Gaussian response. Learning of the model is performed through Bayesian Inference. We develop an efficient MCMC algorithm targeting the posterior distribution of model parameters and missing data (available in GitHub). The motivation for our model lies in the context of Financial Networks, specifically the analysis of transactions between commercial banks. We evaluate the model in synthetic data, as well as our main case study: the network of inter-bank transactions in the Mexican financial system. Accurate predictions are obtained in both cases estimating out-of-sample link incidence and link strength.
Content may be subject to copyright.
Learning of Weighted Multi-layer
Networks via Dynamic Social Spaces,
with Application to Financial
Interbank Transactions
Chris U. Carmona1,3(B
)and Serafin Martinez-Jaramillo2,3
1Department of Statistics, University of Oxford, Oxford, UK
2Center for Latin American Monetary Studies, Mexico City, Mexico
3Banco de Mexico, Mexico City, Mexico
Abstract. We propose a general network model suited for longitudinal
data of multi-layer networks with directed and weighted edges. Our for-
mulation built upon the latent social space representation of networks.
It consists of a hierarchical formulation: deep levels of the model rep-
resent latent coordinates of agents in the social space, evolving in con-
tinuous time via Gaussian Processes; meanwhile, top levels jointly man-
age incidence and strength of interactions by considering a Zero-Inflated
Gaussian response. Learning of the model is performed through Bayesian
Inference. We develop an efficient MCMC algorithm targeting the pos-
terior distribution of model parameters and missing data (available in
GitHub). The motivation for our model lies in the context of Financial
Networks, specifically the analysis of transactions between commercial
banks. We evaluate the model in synthetic data, as well as our main
case study: the network of inter-bank transactions in the Mexican finan-
cial system. Accurate predictions are obtained in both cases estimating
out-of-sample link incidence and link strength.
1 Introduction
In a number of natural, social and economic systems the dynamic interaction
between agents across time can be recorded as longitudinal network-valued data,
with directed and weighted edges. Examples include: migration of people and/or
animals between regions, value of trading between countries, financial transac-
tions between banks, etc. These interactions often occur in multiple layers of con-
nectivity, thus generating multi-layer networks, which should be jointly modeled
for an adequate understanding of the system under study.
The analysis and understanding of Networks have advanced rapidly during
the last few years. Nevertheless, there is still a recognized underdevelopment of
statistical analysis for Dynamic Networks (Crane 2018). The lack of available
models is even more pronounced for dynamic complex systems with weighted,
Springer Nature Switzerland AG 2020
H. Cherifi et al. (Eds.): COMPLEX NETWORKS 2019, SCI 882, pp. 722–735, 2020.
Learning of Weighted Multi-layer Networks 723
directed and multilayered interactions, as most of the seminal advancements in
dynamic network models are centered in binary undirected interactions.
In particular, there exist an important gap in current models for banking
interactions that simultaneously incorporate features such as time-dependancy
across layers, external predictors, and other desirable characteristics that poten-
tially improve the predictive performance of models. There is little to precedent
of statistical inference of the underlying dynamic structure of weighted multi-
layer networks in the context of financial system. Arguably, this may be conse-
quence of two factors: (1) the novelty of methods that can handle the combined
complexity of such systems, and (2) the unavailability of open-access datasets
due to the sensitivity of such data, which in turn deters potential academics
interested on statistical models for financial networks.
A successful implementation of this techniques will uncover critical hidden
structures and dependencies in the financial system, significantly improving the
performance of models for prediction, stress testing, and ultimately, systemic
risk assessment. An adequate understanding of the underlying characteristics of
the complex financial networks provides important insights that translate into
concrete policy insights and policy applications to financial stability and macro-
prudential regulation (Battiston and Martinez-Jaramillo 2018).
1.1 Related Literature: Latent Space Models for Networks
One of the current frontiers in statistical models for networks is being developed
under the Latent Social Space representation. This framework, originally studied
in the seminal work of Hoff et al. (2002), assumes that each agent is positioned
within a latent social space, and the probability of interaction with other agent
depends on the relative distance between the two. The model has been expanded
in recent years to accommodate for more complex features (see Kim et al. (2018a)
for a review).
In Ward et al. (2013), a comprehensive model for longitudinal data of
the world-trade network is introduced. Their model presents valuable features,
such as incorporating network dependencies jointly for both link incidence and
strength, supporting directed relations by considering sender and receiver latent
spaces; as well as including external information as predictors. Nevertheless, this
model is not suited to jointly model multi-layer networks, and the dynamics are
considered in discrete time, which complicates inference when there is uneven
sampling in time.
Durante and Dunson (2014b,2014a) introduced continuous-time dynamics
by considering Gaussian Processes for the evolution of latent coordinates of
agents across time, later expanded in to incorporate locally adaptive dynamics
Durante and Dunson (2016). Durante et al. (2017) introduced a model for multi-
layer networks, the formulation considers one shared latent space to capture
the global structure between agents, and Klayer-specific latent spaces, which
characterises the idiosyncratic structure of each layer. These models, however,
are suitable only for undirected (symmetric) unweighted (binary) edges.
724 C. U. Carmona and S. Martinez-Jaramillo
Sewell and Chen (2015,2016,2017) proposed an alternative formulation for
dynamic networks with directed and weighted edges. Instead of considering two
latent spaces (sender/receiver) as in Ward et al. (2013), they consider one latent
space, together with a set of node-specific parameters r1:nto reflect the social
reach of agents, and two global parameters, βIN and βOUT , to express the impor-
tance of popularity and sociability, respectively. multi-layer networks and con-
tinuous time are not considered in their formulation.
Recently, Linardi et al. (2017) developed a model for inter-bank data based
on Sewell and Chen (2015) model. Their aim is to characterize only the presence
or absence of links between banks in discrete time, in a single layer.
1.2 Our Contribution
We fill a gap in the current literature of statistical network models by combining
features from preceding research into a comprehensive model. Compared with
previous network models, we expanded the inferential tools to allow the inference
and prediction of a more general class of dynamic networks, accommodating for
links with direction, weight, or both. Moreover, the simultaneous incorporation
of additive effects and external covariates had not been discussed before, such
factors reveal important insight about the underlying structure of the network
and its evolution.
The correct implementation of this features together is not trivial. We devel-
oped the DynMultiNet package1based on R and C++. The application of the
model and software developed are not restricted to financial networks, there is
plethora of fields in which our model could be beneficial for the understanding
and prediction of such complex networks.
In the context of financial stability, our work is pioneer. No previous work
has performed statistical inference on the joint network of transactions between
banks. Our probabilistic framework opens the door for future work in anomaly
detection, stress testing and risk management which considers agents interactions
This work is motivated by the study of transactions between commercial
banks within the Mexican Banking System. In Sect. 5we apply the proposed
methodology to real transactions observed in the Mexican Banking System
between 2010 and 2018. Our model achieves accurate in- out-of-sample esti-
mation of probabilities of connection and transaction value.
The model is defined and discussed in Sect. 2, followed by a description of
the MCMC method developed for estimation in Sect. 3.
2 The Model
Consider a dynamic system of Vinteractive agents across time, with activity
recorded as a multi-layer network with Klevels. Let Y(k)(t)={y(k)
ij (t)}∈RV×V
1Available at
Learning of Weighted Multi-layer Networks 725
be the weighted adjacency matrix for the network in layer kat time t, with
ij (t) measuring the interaction from agent ito agent j,fori, j =1,...,V,
Tand k=1,...,K.
The network model consists of a hierarchical model. The observational model
is given by a mixture between a Gaussian distribution and a probability mass
at zero. The first component accounts for the strength of positive interactions
between agents, while the second comprise the pairs with no activity. This is
ij (t)λ(k)
ij (t)∗N(μ(k)
ij (t)
ij (t)] δy(k)
ij (t)(0).(1)
independently for i, j =1,...,V,t=t1,...,t
Tand k=1,...,K. Here, N(μ, σ2)
denotes the Gaussian distribution with mean μand variance σ2and δy(0) the
Dirac measure concentrated at zero.
ij (t) is interpreted as the probability of incidence for an interaction from
agent ito agent jin layer kat time t,andμ(k)
ij (t)istheexpected strength of
such the interaction -if it existed-. σ2
(k)is the variance of all interactions within
the corresponding layer k. We incorporate the latent space representation of
networks through μ(k)
ij (t)andλ(k)
ij (t).
We consider multiple social spaces to deal with the directed and multi-layer
nature of our networks. To accommodate for directed interactions, we follow
Ward et al. (2013) and duplicate the social spaces into the sender and receiver
spaces. Additionally, following Durante et al. (2017), we will use K+ 1 spaces
for a K-layered network: one global space capturing the systemic interaction
between agents, and Klayer-specific spaces that speak for idiosyncratic behavior
for activity of type k.
ij (t)=θ(k)(t)+
sμ,i(t)+pμ,j (t)+
ij (t),
ij (t)=η(k)(t)+
sλ,i(t)+pλ,j (t)+
ij (t).
Let us start by describing the expected strength. Here ui(t)RHis a
dynamic vector of latent coordinates which represents the location of node i
at time twithin the global sender space; similarly, vj(t) is the position of agent j
in the global receiver space.u(k)
j(t) denote the corresponding coordi-
nates within the layer-specific spaces. The baseline processes θ(k)(t)Rcaptures
the average intensity of interactions between all agents in layer k.
Now, sμ,i(t)andpμ,j (t) are additive effects, induced by the sender agent i
and receiver agent j, respectively. Additive effects represent agent i“sociabil-
ity” (out-degree) and agent j“popularity” (in-degree). They capture significant
heterogeneity in activity levels across nodes (Ward et al. 2013;Ho2018;Kim
et al. 2018b).
726 C. U. Carmona and S. Martinez-Jaramillo
External covariates are also a desirable feature that incorporate exogenous
variation to our system (Ward et al. 2013; Durante and Dunson 2014a;Kim
et al. 2018b). We introduce them using a vector of Pedge-specific covariates
ij (t)Rp. We define βμ(t)=(βμ,1(t),...,β
μ,P (t)) Rpand βλ(t)Rp
as the corresponding -dynamic- coefficients for the strength and incidence of
The dynamic of the network is captured by assuming changes in the parame-
ters described before. We adopt smooth trajectories in our formulation introduc-
ing Gaussian Processes (GPs) with squared exponential correlation. C(t, t)=
exp ( tt
δ)2, as the priors for the coordinates,
θ(k)(·)GP (¯
ui,h(·)GP ui,h,C
i,h(·)GP vi,h ,C
i,h (·)GP u(k)
i,h ,C
i,h (·)GP v(k)
i,h ,C
independently for k=1,...,K,i=1,...,V and h=1,...,H, with Cμ(t, t)=
exp ( tt
δμ)2. The parameter and δμcontrol the smoothness of changes across time
for these latent coordinates. The current implementation of our model uses a sin-
gle smoothness parameter for all latent coordinates to ease the MCMC imple-
mentation2. We use a constant (non-zero) mean function for the GPs centering
More elaborate dynamics can be considered for the latent locations. One
option is incorporating non-stationarity through the GPs covariance Ras-
mussen and Williams (2005). Another possibility are Nested Gaussian Processes
(Durante and Dunson 2016) which induce Locally Adaptive trajectories for the
latent factors.
The dimension of the social spaces His fixed. We suggest choosing Hby
optimizing a metric for predictive performance, such as the WAIC Watanabe
(2009) or the Pareto-smoothed approximation Vehtari et al. (2017) to leave-one-
out cross-validation.
The probability of incidence λ(k)
ij (t) is treated in a similar way to the strength
of interaction. However, we need to map the latent similarity measure among
units into the probability space. We use a logistic link for this mapping, obtaining
ij (t)= 1
1 + exp(γ(k)
ij (t)),
ij (t)=η(k)(t)+ai(t)bj(t)+a(k)
We introduced a separate baseline processes η(k)RHand new set of latent
coordinates ai,h,b
i,h ,b
i,h RH. These terms follow similar dynamics as we
2It is possible to optimize for the parameter δfor each agent using Variational Infer-
ence (Tran et al. 2015), which will be explored in future works.
Learning of Weighted Multi-layer Networks 727
defined previously for θ,uand v,
η(k)(·)GP η(k),C
ai,h(·)GP ai,h,C
i,h(·)GP (¯
i,h (·)GP a(k)
i,h ,C
i,h (·)GP (¯
i,h ,C
independently for k=1,...,K,i=1,...,V and h=1,...,H, with Cλ(t, t)=
exp ( tt
We use separate sets of latent coordinates for the incidence and for the
strength of interactions. In case studies datasets, we observed that the empirical
distribution of the log-strength log(y(k)
ij (t) + 1) for existing links (i.e. y0)
is concentrated far from zero, even for pairs with low activity. Using a zero-
truncated gaussian distribution to account for both incidence and strength (as
in Sewell and Chen (2016)) would notably reduce the performance of the model.
We preserve GPs for the dynamics of additive effects and coefficients associ-
ated with external covariates. We use different parameters for the GPs for each
external covariates. In financial networks, agents may react fast to changes in
stock markets levels and slow to changes in macroeconomic variables such as
employment or GDP growth. We have
sμ,i(·)GP sμ,i ,C
λ,i(·)GP sλ,i ,C
pμ,i(·)GP pμ,i,C
λ,i(·)GP pλ,i,C
βμ,l(·)GP (¯
λ,l(·)GP (¯
independently for each agent i=1,...,V and each covariate l=1,...,P.
The formulation of the model allows to straightforwardly handle missing
values, infer missing links and perform forward and backward predictions.
Sampling from missing link is an optional step. Inference of all the latent
parameters does not rely on the imputed values, as they can be integrated out
for the computation of conditional posteriors.
3 Estimation
We adopt a Bayesian approach to learn the parameters in the model. The pos-
terior distribution is computed via a Gibbs sampler. We follow Durante and
Dunson (2014b) and employ a P`olya-Gamma data augmentation strategy Pol-
son et al. (2013) to sample the dynamic terms related to the link incidence
(i.e. η, a, b, a(k),b
λ). Under this approach, every dynamic latent in
the model has a multidimensional Gaussian distribution as conditional poste-
rior. The main complication arise in correctly setting the updating sequence
to maintain conditional independence, and in arranging the associated design
and response matrices. We use Metropolis-Hastings (MH) steps to sample the
variance of the weights σ(k).
728 C. U. Carmona and S. Martinez-Jaramillo
We provide an implementation of the sampler using the R and C++ program-
ming languages3, as well as detailed derivation of the posterior in the author’s
4 Simulation Study
We conduct a simulation study to evaluate the performance of our model in
a synthetic dataset, constructed to mimic a possible generating process in a
financial application4.
Consider a system with V= 15 agents, whose interactions are recorded in a
Weighted multilayer network with K= 2 layers. We assume a total of T=30
observational times, which are randomly (uniform) distributed in the interval
Probabilities of positive interactions are constructed considering two compo-
nents. First, core-periphery static probabilities that differentiate between highly-
connected and dissociated nodes, as shown in heat maps on the left side of
Fig. 1(top layer 1, bottom layer 2). Second, dynamic probabilities that emu-
late a scenario of financial crisis happening at time t= 10, displayed on Fig. 1.
Two of these trajectories affect the probability λ(k)
ij (t), one for each layer: a
sharp increase in connectivity before the crisis, followed by a prolonged decrease
(top-left); a systemic decrease of trust that start few periods before the crisis
The expected weights of the links are created through an inverse logis-
tic transformation of the static probabilities, then adding a fixed value for
each layer, and finally including two additional trajectories. The trajecto-
ries here assume that strength of connection increases/decreases between cen-
tral/periphery agents after the crisis (bottom-left/bottom-right).
Finally, both expected weights and probabilities were perturbed considering
agent-specific trajectories during the period.
For inference, we choose a value H= 3 for the dimension of the latent spaces
and a value δ= 20 for the smoothing parameter. We ran 2500 Gibbs iterations,
and discarded the first 500 as warm-up period. We verified good mixing of the
parameters in the observational equations.
In Fig. 2we compare the generative vs. estimated probabilities of connection
for three relevant times: before, at, and after the crisis (t=0.3,9.1,24.6 respec-
tively). The structure of the network is learned adequately. Figure 3shows an
overall comparison of the estimated probabilities and weights across all agents,
layers and times. The panels compare horizontally three types of estimation: in-
sample, imputation, and projective forecasting. Top panels correspond to proba-
bilities, bottom panels to weights. The model effectively captures the underlying
structure of the network in all periods, but it perform much better when more
information is available, as expected.
3Available at
4Code to replicates this experiment is available at the author’s website.
Learning of Weighted Multi-layer Networks 729
Fig. 1. Base probabilities and systemic trends in synthetic data. Left panel shows base
probabilities for the V= 15 agents and two K= 2 layers. Right panel show the
systemic trends that modify μ(k)
ij (t)andμ(k)
ij (t) through time for all agents. The points
corresponding values at observational times. Shaded areas display periods with missing
Fig. 2. Comparison of real vs. estimated probabilities. We show the probability of con-
nection λ(k)
ij (t) for three times (t=0.3,9.1,24.6) and the K= 2 layers. The left panel
shows the original generative probabilities, while the right panel show the estimation
given by the posterior mean.
Lastly, Fig. 4displays how the model captures the dynamics of the network.
For the pair of agents (i=1,j = 7), we show the evolution of probability of
connection and link strength. The smooth dynamic of the latent elements in the
model effectively captures changes in the adjacency matrices.
730 C. U. Carmona and S. Martinez-Jaramillo
Fig. 3. Overall comparison of real vs. estimated probabilities and weights. We dis-
play the prediction accuracy for probability λ(k)
ij (t) (top) and mean strength μ(k)
ij (t)
(bottom). We compare the prediction under three conditions: Period with observations
(left), period with missing data within sample 15 <t<18 (center), and forecasted
values for the future t>23 (right).
Fig. 4. Dynamics of probability of connection λ(k)
ij (t) (left) and expected weight μ(k)
ij (t)
(right) for the pair of agents (i=1,j =7)inlayerk= 1. Red lines show the posterior
mean for both quantities, together with their 95% and 75% credibility intervals. Blue
lines are the true generative quantities. Blue circles show the network data: solid for
observed, empty for missing data.
5 Case Study
5.1 Interbank Activity in the Mexican Financial System
The data used for this case study is derived from a database on exposures built
and operated by the Central Bank of Mexico (Banxico) with the specific pur-
pose of studying contagion and systemic risk. Banxico gathers information using
daily, weekly, and monthly regulatory reports. These reports contain every single
funding transaction between most financial institutions in the Mexican Financial
Learning of Weighted Multi-layer Networks 731
System, on a daily basis, in local and foreign currency. For empirical network
studies using this data see Poledna et al. (2015); Molina-Borboa et al. (2015);
de la Concha et al. (2018).
We aim to characterise the underlying probabilistic structure that dictates
interactions between banks across markets. The results of the inferential pro-
cess will provide enlightening information about the underlying connectivity of
the banking system. We focus our analysis on interbank transaction within the
Mexican financial system in three layers:
OCIMN + OCIME: Unsecured lending. This network is built using the infor-
mation from the regulatory report which contain the unsecured loans in
domestic currency (OCIMN) and in foreign currency (OCIME). A link ij
is the sum of all the transacted loans granted by ito jin any currency during
the time period of aggregation (daily or monthly).
REPO: Repurchase agreement transactions. This network is built using the
information from the regulatory report on Repo transactions. A link ij is
the sum of all the Repos transacted during the time period of aggregation,
regardless of the collateral, currency and maturity.
CVT: Purchase and sale of securities, bilateral transactions. This network is
built with the information from the regulatory report on the bond market
activity by banks. A link ij represents the total amount of all the securi-
ties sold by ito jregardless of the type of securities and the terms of such
We discuss here the results from two main experiments. We analyze financial
data from V= 46 banks in the layers described above. We used aggregated
monthly data during two periods of analysis to train the model:
2008-01-01 to 2009-12-31 (two years observed)
2008-01-01 to 2010-12-31 (three years observed)
The predictive horizon was set as 1 year forward in both cases. For inference,
we set δ= 24, H=5.
There is a clear structure captured by the model, which allows to easily
identify the centrality and relevance of each actor in the system. The estimated
propensity of connection between agents in the system for a given month is
illustrated in Fig. 5. Each bank is represented by a circle in this graph, whereas
the distance between them and the width of line linking them represents the
probability of being connected.
The dynamic of the interaction between banks is also learned by the model.
In Fig. 6, the evolution of the relation between two banks is depicted. The left
panel shows the probability of connection between them as a function of time.
We see that there is a clear gradual increase in unsecured transactions from
bank “11” to bank “1” during 2019 (time 13 to 24). The right panel shows the
expected volume for the transaction (conditional on its occurrence). The blue
points in both graphs correspond to the observed outcome (0/1 for link existence
in the left and a positive weight for the transaction volume in the right.).
732 C. U. Carmona and S. Martinez-Jaramillo
Fig. 5. Force–directed visualization of unsecured market for a (randomly) selected day,
estimated by the model. The width of the edges are proportional to the probability
of connection between two agents on that day. The size of the nodes is given by their
aggregated in-strength.
Fig. 6. For a (random) pair of banks (i=11,j = 1), we show the dynamic of the
strength of unsecured transactions from bank ito bank jacross time (monthly activ-
ity, time = 0 corresponds to Jan-2008). On the left, the posterior distribution for the
probability of connection, Mean ˆpin red, and the observed activity represented by the
blue dots (0/1). On the right, The expected amount for the transaction (conditional
on being connected), and the actual amount also represented by the blue dots. All
quantities are estimated simultaneously within a comprehensive Bayesian model.
We evaluate predictive accuracy of edge formation for each layer as measured
by the ROC curve and the associated area under the ROC Curve, AUC. Out-
of-sample performance was obtained by predicting 12 months following after the
training period and comparing to the observed (left-out) values. In Fig. 7we
show the Area-Under-the-Curve for predictions made for both in-sample and
out-of-sample periods. The performance is outstanding, reaching almost 1 for
in-sample edges; slowly decaying for out-of-sample data, but still with values
above 0.85 for all layers.
Regarding the prediction of weights, in Fig.8we compare observed vs pre-
dicted weights for the three layers. The left plot corresponds to in-sample obser-
vations, reaching a strong correlation of 0.89; whereas the right plot corresponds
to predicting the first six months after the training data, showing a correlation
of 0.77.
Learning of Weighted Multi-layer Networks 733
Fig. 7. Assessing model accuracy. Area Under the ROC curve.
Fig. 8. Assessing model accuracy. Comparison of observed vs predicted link weights.
6 Conclusions
This work provides a comprehensive statistical model that can be useful to cap-
ture in a very broad sense complex aspects of interconnectedness in weighted
multilayer networks. There very limited number of statistical models which com-
prise weighted and directed multiplex structures, something which is a crucial
element in financial networks applications, in particular those related to systemic
risk analysis. The proposed model proved to be effective in capturing important
aspects of the complexity of financial networks.
The possible implications in the field of stress testing and systemic risk mea-
surement are many and important. Until now, most of the studies on financial
networks rely on limited information to build financial networks (one layer) or
single shots (limiting the analysis of the dynamics of such structures).
The model and the software developed usefulness is not restricted to financial
networks, there is plethora of fields in which it could be useful and benefit from
all these novel techniques which were applied in this work.
734 C. U. Carmona and S. Martinez-Jaramillo
Battiston, S., Martinez-Jaramillo, S.: Financial networks and stress testing: challenges
and new research avenues for systemic risk analysis and financial stability implica-
tions. J. Financ. Stab. 35, 6–16 (2018)
Crane, H.: Probabilistic Foundations of Statistical Network Analysis, 1st edn. Chapman
and Hall/CRC (2018)
de la Concha, A., Martinez-Jaramillo, S., Carmona, C.: Multiplex financial networks:
revealing the level of interconnectedness in the banking system. In: Complex Net-
works & Their Applications VI, pp. 1135–1148. Springer (2018)
Durante, D., Dunson, D.B.: Bayesian dynamic financial networks with time-varying
predictors. Stat. Probab. Lett. 93, 19–26 (2014a)
Durante, D., Dunson, D.B.: Nonparametric Bayes dynamic modelling of relational data.
Biometrika 101(4), 883–898 (2014b)
Durante, D., Dunson, D.B.: Locally adaptive dynamic networks. Ann. Appl. Stat.
10(4), 2203–2232 (2016)
Durante, D., Mukherjee, N., Steorts, R.C.: Bayesian learning of dynamic multilayer
networks. J. Mach. Learn. Res. 18, 1–29 (2017)
Hoff, P.D.: Additive and multiplicative effects network models (2018)
Hoff, P.D., Raftery, A.E., Handcock, M.S.: Latent space approaches to social network
analysis.J.Am.Stat.Assoc.97(460), 1090–1098 (2002)
Kim, B., Lee, K.H., Xue, L., Niu, X.: A review of dynamic network models with latent
variables. Stat. Surv. 12, 105–135 (2018a)
Kim, B., Niu, X., Hunter, D.R., Cao, X.: A dynamic additive and multiplicative effects
model with application to the united nations voting behaviors (2018b)
Linardi, F., Diks, C.G.H., van der Leij, M., Lazier, I.: Dynamic interbank network
analysis using latent space models. SSRN Electron. J. (2017)
Molina-Borboa, J., Mart´ınez-Jaramillo, S., Lopez-Gallo, F.: A multiplex network analy-
sis of the Mexican banking system: link persistence, overlap. J. Netw. Theory Finance
1(1), 99–138 (2015)
Poledna, S., Molina-Borboa, J.L., Mart´ınez-Jaramillo, S., van der Leij, M., Thurner,
S.: The multi-layer network nature of systemic risk and its implications for the costs
of financial crises. J. Financ. Stab. 20, 70–81 (2015)
Polson, N.G., Scott, J.G., Windle, J.: Bayesian inference for logistic models using P´olya-
Gamma latent variables. J. Am. Stat. Assoc. 108(504), 1339–1349 (2013)
Rasmussen, C.E., Williams, C.K.I.: Gaussian Processes for Machine Learning. The
MIT Press (2005)
Sewell, D.K., Chen, Y.: Latent space models for dynamic networks. J. Am. Stat. Assoc.
110(512), 1646–1657 (2015)
Sewell, D.K., Chen, Y.: Latent space models for dynamic networks with weighted edges.
Soc. Netw. 44, 105–116 (2016)
Sewell, D.K., Chen, Y.: Latent space approaches to community detection in dynamic
networks. Bayesian Anal. 12(2), 351–377 (2017)
Tran, D., Blei, D., Airoldi, E.M.: Copula variational inference. In: Cortes, C., Lawrence,
N.D., Lee, D.D., Sugiyama, M., Garnett, R. (eds.) Advances in Neural Information
Processing Systems, vol. 28, pp. 3564–3572. Curran Associates, Inc. (2015)
Learning of Weighted Multi-layer Networks 735
Vehtari, A., Gelman, A., Gabry, J.: Practical Bayesian model evaluation using leave-
one-out cross-validation and WAIC. Stat. Comput. 27(5), 1413–1432 (2017)
Ward, M.D., Ahlquist, J.S., Rozenas, A.: Gravity’s rainbow: a dynamic latent space
model for the world trade network. Netw. Sci. 1(01), 95–118 (2013)
Watanabe, S.: Algebraic Geometry and Statistical Learning Theory. Cambridge Mono-
graphs on Applied and Computational Mathematics. Cambridge University Press
... On a more recent strand of the literature on financial networks, the financial system and the banking system in particular is put into the lights of multilayer networks as in Montagna and Kok (2013) , Poledna et al. (2015) , Bravo-Benitez et al. (2016) , Bookstaber and Kennet (2016) , Concha et al. (2018) , Aldasoro and Alves (2018) , Carmona and Martinez-Jaramillo (2020) , Poledna et al. (2021) and Cuba et al. (2021) . Indeed, banks and other financial intermediaries interact in many different markets at the same time and those interactions serve different purposes and might create different types of risks. ...
Full-text available
Systemic risk analysis has become a very important undertaking in most central banks after the Global Financial Crisis (GFC). This paper describes the Colombian credit system of banks and firms as a bipartite network of lenders and borrowers. To such network, we apply a spectral method to identify the most central actors, and a variant of the DebtRank algorithm to identify the banks and firms that would be the most vulnerable to shocks in the system, and the most impactful in propagating them. We perform our analysis with a multi-layer approach, analysing networks of loans in the Commercial, Housing, and Microcredit domain. Our analyses reveal a rich and heterogeneous systemic risk profile across the Colombian credit system, and highlight the presence of considerable network effects that would contribute to shape the propagation of shocks from the real economy to the banking system.
Conference Paper
Full-text available
Complex networks models have been useful for the study of systemic risk; however, most of the studies have ignored the true level of interconnectedness in the financial system; in this work we show the missing part on the study of interconnectedness. Furthermore, complexity in modern financial systems has been also an important subject of study. However, we still lack the appropriate metrics to describe such complexity and interconnectedness; moreover, the data available in order to describe and study them is still scarce. In addition, most of the focus on the subject of interconnectedness has been on a single type of network: interbank (exposures) networks. We use a comprehensive set of market interactions that include transactions in the securities markets, payment system flows, interbank loans, cross holding of securities, foreign exchange exposures and derivatives exposures. This the first attempt, to the best of our knowledge, to describe so comprehensively the complexity and interconnectedness in a banking system. We are able to identify the most important institutions in the whole structure in term of their connectedness, the most relevant layer (in structural terms) of the multiplex and the community structure of the Mexican banking system which can be seen as a generalization of the well-known Core-Periphery model.
Full-text available
Our focus is on realistically modeling and forecasting dynamic networks of face-to-face contacts among individuals. Important as- pects of such data that lead to problems with current methods in- clude the tendency of the contacts to move between periods of slow and rapid changes, and the dynamic heterogeneity in the actors' con- nectivity behaviors. Motivated by this application, we develop a novel method for Locally Adaptive DYnamic (LADY) network inference. The proposed model relies on a dynamic latent space representation in which each actor's position evolves in time via stochastic di�eren- tial equations. Using a state space representation for these stochastic processes and P�olya-gamma data augmentation, we develop an e�- cient MCMC algorithm for posterior inference along with tractable procedures for online updating and forecasting of future networks.We evaluate performance in simulation studies, and consider an applica- tion to face-to-face contacts among individuals in a primary school.
Full-text available
Leave-one-out cross-validation (LOO) and the widely applicable information criterion (WAIC) are methods for estimating pointwise out-of-sample prediction accuracy from a fitted Bayesian model using the log-likelihood evaluated at the posterior simulations of the parameter values. LOO and WAIC have various advantages over simpler estimates of predictive error such as AIC and DIC but are less used in practice because they involve additional computational steps. Here we lay out fast and stable computations for LOO and WAIC that can be performed using existing simulation draws. We introduce an efficient computation of LOO using Pareto-smoothed importance sampling (PSIS), a new procedure for regularizing importance weights. Although WAIC is asymptotically equal to LOO, we demonstrate that PSIS-LOO is more robust in the finite case with weak priors or influential observations. As a byproduct of our calculations, we also obtain approximate standard errors for estimated predictive errors and for comparison of predictive errors between two models. We implement the computations in an R package called loo and demonstrate using models fit with the Bayesian inference package Stan.
Full-text available
A plethora of networks is being collected in a growing number of fields, including disease transmission, international relations, social interactions, and others. As data streams continue to grow, the complexity associated with these highly multidimensional connectivity data presents new challenges. In this paper, we focus on time-varying interconnections among a set of actors in multiple contexts, called layers. Current literature lacks flexible statistical models for dynamic multilayer networks, which can enhance quality in inference and prediction by efficiently borrowing information within each network, across time, and between layers. Motivated by this gap, we develop a Bayesian nonparametric model leveraging latent space representations. Our formulation characterizes the edge probabilities as a function of shared and layer-specific actors positions in a latent space, with these positions changing in time via Gaussian processes. This representation facilitates dimensionality reduction and incorporates different sources of information in the observed data. In addition, we obtain tractable procedures for posterior computation, inference, and prediction. We provide theoretical results on the flexibility of our model. Our methods are tested on simulations and infection studies monitoring dynamic face-to-face contacts among individuals in multiple days, where we perform better than current methods in inference and prediction.
Full-text available
Embedding dyadic data into a latent space has long been a popular approach to modeling networks of all kinds. While clustering has been done using this approach for static networks, this paper gives two methods of community detection within dynamic network data, building upon the distance and projection models previously proposed in the literature. Our proposed approaches capture the time-varying aspect of the data, can model directed or undirected edges, inherently incorporate transitivity and account for each actor's individual propensity to form edges. We provide Bayesian estimation algorithms, and apply these methods to a ranked dynamic friendship network and world export/import data.
Network models, stress testing methods and early warning systems are attracting growing interest both among scholars and practitioners. In this short paper, we illustrate some examples of the insights they have to offer both in terms of new fundamental scientific understanding and in terms of concrete applications to the policy area of financial stability and macro-prudential policy. Finally, we discuss some new research pathways that the field could engage in order to address the challenging questions still open.
In this paper, we introduce a statistical regression model for discrete-time networks that are correlated over time. Our model is a dynamic version of a Gaussian additive and multiplicative effects (DAME) model which extends the latent factor network model of Hoff (2009) and the additive and multiplicative effects model of Minhas et al. (2016a), by incorporating the temporal correlation structure into the prior specifications of the parameters. The temporal evolution of the network is modeled through a Gaussian process (GP) as in Durante and Dunson (2013), where we estimate the unknown covariance structure from the dataset. We analyze the United Nations General Assembly voting data from 1983 to 2014 (Voeten et al., 2016) and show the effectiveness of our model at inferring the dyadic dependence structure among the international voting behaviors as well as allowing for a varying number of nodes over time. Overall, the DAME model shows significantly better fit to the dataset compared to alternative approaches. Moreover, after controlling for other dyadic covariates such as geographic distances and bilateral trade between countries, the model-estimated additive effects, multiplicative effects, and their movements reveal interesting and meaningful foreign policy positions and alliances of various countries.
We present a selective review on statistical modeling of dynamic networks. We focus on models with latent variables, specifically, the latent space models and the latent class models (or stochastic blockmodels), which investigate both the observed features and the unobserved structure of networks. We begin with an overview of the static models and then we introduce the dynamic extensions. For each dynamic model, we also discuss its applications that have been studied in the literature, with the data source listed in Appendix. Based upon the review, we summarize a list of open problems and challenges in dynamic network modeling with latent variables.