Bayesys datasets

Various datasets from the Bayesys repository.

Size: 6 groups of datasets with each up to 16 experimentally generated from the bayesian network with the number of observation 100,1000,…100000. Ground truth is given

Number of features: 6 - over 1000

Ground truth: Yes

Type of Graph: Directed graph

Six discrete BN case studies are used to generate data. The first three of them represent well-established examples from the BN structure learning literature, whereas the other three represent new cases and are based on recent BN real-world applications. Specifically,

Asia: A small toy network for diagnosing patients at a clinic;

Alarm: A medium-sized network based on an alarm message system for patient monitoring;

Pathfinder: A very large network that was designed to assist surgical pathologists with the diagnosis of lymph-node diseases;

Sports: A small BN that combines football team ratings with various team performance statistics to predict a series of match outcomes;

ForMed: A large BN that captures the risk of violent reoffending of mentally ill prisoners, along with multiple interventions for managing this risk;

Property: A medium BN that assesses investment decisions in the UK property market.

Data generated with noise:

Synthetic datasets - noise

Experiment No.
Experiment
Notes

1
N
No noise

2
M5
Missing data (5%)

3
M10
Missing data (10%)

4
I5
Incorrect data (5%)

5
I10
Incorrect data (10%)

6
S5
Merged states data (5%)

7
S10
Merged states data (10%)

8
L5
Latent confounders (5%)

9
L10
Latent confounders (10%)

10
cMI
M5 and I5

11
cMS
M5 and S5

12
cML
M5 and L5

13
cIS
I5 and S5

14
cIL
I5 and L5

15
cSL
S5 and L5

16
cMISL
M5, I5, S5 and L5

More information about the datasets is contained in the dataset_description.html files.