Genetic Algorithms for Tractable Bayesian Network Consensus

This repository contains the source code for the paper "Genetic Algorithms for Tractable Bayesian Network Consensus: A Unified Treewidth-Constrained Framework".

Overview

Bayesian network structural fusion combines locally learned networks into a consensus model, which is central to federated and distributed learning. However, the unrestricted fusion creates dense graphs whose high treewidth makes exact inference intractable.

This repository implements a unified framework that tackles two complementary problems under a treewidth constraint:

1. Fusion approximation (P1): Approximating the unrestricted fusion under a treewidth bound.
2. Structural consensus (P2): Finding the consensus network most representative of the inputs under the same constraint.

Four genetic algorithm variants parameterize the search space differently:

• GA-$\mathcal{E}_a$: Pruning arcs from the post-fusion graph.
• GA-$\mathcal{E}_c$: Pruning unique arcs from the input networks before fusion.
• GA-$\mathcal{E}_b$: Pruning arcs with an independent decision per network.
• GA-$\mathcal{E}_b^{MC}$: A hybrid variant that uses a min-cut-based greedy method as a warm start to seed the GA-$\mathcal{E}_b$ population.

Prerequisites

• Java 17
• Maven

Build Instructions

To compile the project and build the executable JAR with dependencies, run:

mvn clean package

This will generate a JAR file in the target/ directory (e.g., bnfusion-1.0-jar-with-dependencies.jar).

Running the Experiments

The main entry point for running the experiments is the consensusBN.Experiments.ExperimentsJournal class. This class runs all the algorithms (greedy and genetic variants) for a given network and parameters at different treewidth values.

Run the compiled JAR:

java -cp target/bnfusion-1.0-jar-with-dependencies.jar consensusBN.Experiments.ExperimentsJournal [index] [params_file]

Or just run without arguments to execute a quick local test (by default it will run alarm with 10 clients and output parameters):

java -cp target/bnfusion-1.0-jar-with-dependencies.jar consensusBN.Experiments.ExperimentsJournal

Parameters File

The program reads a text file line by line using the index (0-indexed) to select the specific run's configuration. Each line should be space-separated with the following parameters:

1. net: Name of the network (e.g., alarm.0)
2. nClients: Number of clients/input DAGs (e.g., 10)
3. popSize: Population size for the GA (e.g., 100)
4. nIterations: Number of generations for the GA (e.g., 100)
5. twLimit: Minimum Treewidth limit setting (e.g., 2)
6. seed: Random seed for reproducibility (e.g., 0)
7. useGES (Optional): true to learn input DAGs using GES from data, false to generate perturbations (default: false).
8. optimizeSMHD (Optional): true to use SMHD fitness, false to use FSim fitness (default: true).
9. optimizeAgainstOriginals (Optional): true to measure against input DAGs (P2), false to measure against G+ (P1) (default: true).
10. useGreedyWarmstart (Optional): true to inject greedy result into population, false for random init (default: true).

Output

The results are saved as .csv files inside the results/Server/ folder, and the convergence curves for the genetic algorithms are stored in the results/Server/convergence/ subfolder.