A 'Shiny' web application for creating interactive Bayesian Network models, learning the structure and parameters of Bayesian networks, and utilities for classical network analysis.
BayesianNetwork is a Shiny web application for Bayesian network modeling and analysis, powered by the excellent bnlearn and networkD3 packages. This app is a more general version of the RiskNetwork web app. To learn more about our project, check out this publication. For even more background, see this publication.
To install BayesianNetwork in R:
Or to install the latest developmental version:
To launch the app:
Or to access the app through a browser, visit paulgovan.shinyapps.io/BayesianNetwork.
Launching the app brings up the Dashboard tab. The Dashboard tab is basically a landing page that gives a brief introduction to the app and includes two value boxes, one each for the number of nodes and arcs in the network. The following figure shows the basic Dashboard tab.
BayesianNetwork comes with a number of simulated and "real world" data sets. This example will use the "Sample Discrete Network", which is the selected network by default.
Click Structure in the sidepanel to begin learning the network from the data. The Bayesian network is automatically displayed in the Bayesian Network box.
In order to learn the structure of a network for a given data set, upload the data set in csv format using The Network Input box. Again, this example uses the "Sample Discrete Network", which should already be loaded.
Select a learning algorithm from the Structural Learning box. The classes of available structural learning algorithms include:
To view the network score, select a score function from the The Network Score box.
"Sample Discrete Network" contains six discrete variables, stored as factors with either 2 or 3 levels. The structure of this simple Bayesian network can be learned using the grow-shrink algorithm, which is the selected algorithm by default.
Try different combinations of structural learning algorithms and score functions in order to see the effect (if any) on the resulting Bayesian network.
Select the grow-shrink algorithm once again and then click Parameters in the sidepanel in order to learn the parameters of the network. The selected paramaters are automically displayed in the Network Paramaters box.
Select a learning algorithm from the Parameter Learning box. This app supports both maximum-likelihood and Bayesian estimation of the parameters. Note that Bayesian parameter learning is currently only implemented for discrete data sets. Then select the type of chart to display in the Paramter Infographic box and, for the discrete case, choose the preferred node. For example, the selected node A is a discrete node with three levels: a, b, and c.
Click Measures in the sidepanel to bring up a number of tools for classic network analysis. The Measures tab has a number of node and network measures. The node measures include:
Select a node measure in the Node Control box and the result will be displayed in the Node Measure box.
The Network Control box contains different options for displaying hierarchical clusters/dendograms for the network. Select the type of dendogram to display (row, column, both, or none) and the resulting dendogram(s) will be displayed along with the adjacency matrix in the Network Measure box.
Finally, click Simulation in the sidepanel in order to simulate data from the network for future use. Simply enter the sample size N in the Network Simulation box and click Download in order to download a random sample of data from the Bayesian network.
This package is very much still in development and mutiple improvements are planned, including:
Please let us know if there are things you would like to see added (or problems with the app!) by opening up an issue using the GitHub issue tracker at https://github.com/paulgovan/BayesianNetwork/issues
Contributions are welcome by sending a pull request
BayesianNetwork is licensed under the Apache licence. © Paul Govan (2015)