Hidden variables in a Dynamic Bayesian Network identify ecosystem level change

Abstract

Ecosystems are known to change in terms of their structure and functioning over time. Modelling this change is a challenge, however, as data are scarce, and models often assume that the relationships between ecosystem components are invariable over time. Dynamic Bayesian Networks (DBN) with hidden variables have been proposed as a method to overcome this challenge, as the hidden variables can capture the unobserved processes. In this paper, we fit a series of DBNs with different hidden variable structures to a system known to have undergone a major structural change, i.e., the Baltic Sea food web. The exact setup of the hidden variables did not considerably affect the result, and the hidden variables picked up a pattern that agrees with previous research on the system dynamics.

Using Bayesian Networks to Investigate the Influence of Subseasonal Arctic Variability on Midlatitude North Atlantic Circulation

“…Results were robust to the choice of alpha between 0.01 and 0.05. In the case of experiments 2 and 3, a hidden variable with links to all variables and to itself through time (autoregressively to the next time step) is coded into the structure, as in Uusitalo et al (2018). Next, any bidirectional links (which would lead to cyclical structures) are removed from the learned structure by necessity as DBNs require acyclical graphs; DAGs by definition cannot include loops (Scutari 2010; Scutari and Denis 2014).…”

“…Given the challenges associated with identifying AA-midlatitude linkages due to noisy internal dynamics and the time-constrained nature of AA processes, potential improvements in model accuracy make hidden variables worth consideration. Graphical models with hidden variables inferred from observed data have been largely untouched in climate science studies, but their capabilities have been explored and proven in ecological system analyses (Trifonova et al 2015(Trifonova et al , 2017(Trifonova et al , 2019Uusitalo et al 2018).…”

Review of trophic models for the Baltic Sea

“…Dynamic Bayesian Networks (DBN) with hidden variables have been proposed as a method to overcome this challenge, as the hidden variables can capture the unobserved processes. In Uusitalo et al (2018), a series of DBNs with different hidden variable structures was fit to the Baltic Sea food web. The exact setup of the hidden variables did not considerably affect the result, and the hidden variables picked up a pattern that agrees with previous research on the system dynamics.…”

A Comparison of Machine-Learning Methods to Select Socioeconomic Indicators in Cultural Landscapes

“…BNs provide a well-founded approach for dealing with complex systems [44]. The graphical representation provided by BNs makes them a transparent tool, since the different parts of the ecosystem can be seen as nodes that interact with other nodes in a network, and the relationship between them can be mathematically modeled [4,21]. BNs provide not only a numeric prediction of the response but also a full specification of the posterior probability distribution.…”

“…In particular, the machine-learning field has been gaining popularity among the environmental modelers due to the development of sound methods and algorithms able to deal with complex systems without making assumptions of linearity.Machine-learning [13,14] techniques are now fundamental tools in many research areas. This field comprises a large amount of mathematical and statistical methods that have proven to be useful in a vast number of applications, such as animal behavior [15], biological invasions [16], species distribution [17][18][19], food webs [20,21], ecosystem services [4,22], air pollution [23-25], groundwater pollution [26][27][28], surfacewater pollution [29][30][31], or land-cover classification [32][33][34]. The research in these topics has become highly active in the few last years.…”