Unlike Unsupervised Structural Learning, the objective of Supervised Learning is not to find the best representation of the joint probability distribution sampled by the particles described in the data set, but rather to find the best probabilistic characterization of the Target Node.
Among the available Supervised Learning algorithms offered in BayesiaLab, the Markov Blanket Learning algorithms are the most advanced.
Their main objective is to efficiently identify the subset of nodes, the Markov Blanket, that makes the Target node conditionally independent of all the other nodes.
The Markov Blanket of the Target node is defined by:
- its Parents: allow to cut the information flow coming from the indirect ancestors,
- its Children: allow to cut the information flow coming from the indirect descendants,
- its Spouses: allow to cut the conditional information flow coming from the ancestors and descendants of the spouses when the Children are observed;
There are 3 algorithms Markov Blanket Learning algorithms:
- Markov Blanket (MB): return the graphs where the Target is connected with the identified Markov Blanket - can be used for Variable Selection;
- Augmented Markov Blanket (AMB): MB plus unsupervised learning to represent the dependencies between the selected nodes - allows to improve the quality of the model,
- Minimal Augmented Markov Blanket (MAMB): Unsupervised Structural Learning algorithm applied on the nodes selected by MB, and extraction of the Markov Blanket - the Markov Blanket returned by MB may contain the ancestors and descendants that have more than one path to the Target (can be useful for prediction). This algorithm is therefore useful knowledge discovery as it may return a Markov Blanket that is closer to the actual one.