TIBER

Federated bayesian network algorithm for horizontally partitioned data.

Usage

The implementation for TIBER provides a federated approach to build a Bayesian network. Using vantage6 as the infrastructure makes it possible to run the algorithm at each center using the following interface:

    client$set.task.image(
        'tiber/docker-image:x.y.z',
        task.name="bayesian"
    )

    client$use.master.container <- TRUE
    config <- list(
        # arc_strength_args = list(
        #     algorithm = "hc",
        #     R = 300
        #     algorithm.args = list(score = "bde", restart = 5, perturb = 5)
        # ),
        # weighted_strength = 0.3,
        # val_org_id = c(),
        # exclude = c('PT')
    )
    column_to_predict <- 'PN'
    result <- client$call('bayesian', column_to_predict, config)

Calling the main method (bayesian) will build the bayesian network using the default configurations:

• algorithm - the structure learning algorithm used: 'hc'
• R - the number of bootstrap replicates: 400
• weighted_strength - threshold to select the network arcs based on the aggregated scoring: 0.2

These configurations can be overridden using the config list and providing it when calling the vantage6 client. Additionally, it's also possible to configure other parameters:

• val_org_id - explicitly provide the organization(s) id to use as validation. In case the argument isn't provided, one random organization from the collaboration will be chosen. In case an empty vector is provided, no validation will be performed and all organizations will be used for training.
• exclude - columns from the dataset to be ignore when creating and training the network

For a complete overview, check the examples 'run.R' or 'run-tiber.R'.

Packaging the algorithm

This repository contains the code to build the package and create the docker algorithm image for vantage6. To build (and push) the docker image:

• cd to the src directory;
• execute the following command to build the docker image: make docker-build
• execute the following command to build and push the docker image: make docker

Testing

To run the algorithm, use the run.R script. This requires you to have the R version of vantage6 installed.

Mock data

The SyntheticPooledSet.csv file contains data to use when running the algorithm. Each node should contain data from one of the countries from the trial column.