Mixture Model

class pyagrum.bnmixture.BNMixture

A mixture of Bayesian networks where each network carries a positive weight.

The reference BN (refBN) is the BN with the highest weight; it is recomputed by updateRef(). It serves as the structural template for visualisation and variable look-ups, but plays no special role during inference: BNMixtureInference returns the weight-averaged posterior over all BNs in the mixture.

Notes

This is an experimental model.

BN(name)

Parameters: name (str) – Name of the variable.
Returns: A copy of the BN with name name in the model.
Return type: BayesNet
Raises: pyagrum.NotFound – If the given name doesn’t correspond to the name of a BN in the model.

BNs()

Returns: A list containing a copy of all BNs in the model.
Return type: list[BayesNet]

add(name, bn, w=1)

Adds a BN to the model. If the model doesn’t have a reference BN when trying to add an element, the BN (before adding new element) with highest weight becomes the new reference.

Parameters:
- name (str) – Name of the BN to add.
- bn (BayesNet) – BN to add.
- w (float) – Weight of the BN.
Raises:
- pyagrum.InvalidArgument – If the weight is negative.
- pyagrum.InvalidArgument – If the names of the variables in the BN to add are differents from the one in the reference BN.
- pyagrum.InvalidArgument – If the variables in the BN to add are differents from the one in the reference BN.
- pyagrum.InvalidArgument – If the name in argument is the same as the reference BN’s name.
- pyagrum.ArgumentError – If the name in argument already exists for a different BN in the model.
Return type: None

existsArc(a, b)

Counts the number of time arc a -> b appears among all BNs in the model.

Parameters:
- a (str | int) – Tail of the arc.
- b (str | int) – Head of the arc.
Returns: The number of time arc a -> b appears.
Return type: int

isNormalized()

Checks if the model is normalized (the sum of the weights equals 1).

Return type: bool

isValid()

Checks if all the weights are equal to 0. Valid if sum of the weights is not 0.

Returns: True if weights are valid. False otherwise.
Return type: bool

names()

Returns: The list of names of the BNs in the model (reference BN not included).
Return type: list[str]

normalize()

Normalizes the weights.

Raises: pyagrum.InvalidArgument – If all weights are zero (cannot normalize).
Return type: None

property refBN : BayesNet | None

Read-only access to the reference BN.

property refName : str

Read-only access to the reference BN name.

remove(name)

Removes a BN from the model.

Parameters: name (str) – Name of the BN to remove.
Raises: pyagrum.NotFound – If the given name doesn’t correspond to the name of a BN in the model.
Return type: None

setWeight(name, w)

Changes the weight of a BN in the model.

Parameters:
- name (str) – Name of the BN to modify.
- w (float) – Value of the new weight.
Raises:
- pyagrum.InvalidArgument – If the weight is negative.
- pyagrum.NotFound – If the given name doesn’t correspond to the name of a BN in the model.
Return type: None

size()

Returns: The number of BNs in the model (reference BN not included).
Return type: int

updateRef()

Updates the reference BN. The new reference BN is the one with maximum weight.

Return type: None

variable(name)

Parameters: name (str) – Name of the variable.
Returns: The corresponding variable.
Return type: DiscreteVariable

weight(name)

Parameters: name (str) – Name of the BN.
Returns: The weight of the BN with name name.
Return type: float
Raises: pyagrum.NotFound – If the given name doesn’t correspond to the name of a BN in the model.

weights()

Returns: The weights of all the BNs in the model.
Return type: dict[str, float]

zeroBNs()

Returns: The names of the BNs in the model that have weight with value 0.
Return type: set[str]

class pyagrum.bnmixture.BootstrapMixture(name, bn)

A mixture of Bayesian networks built by Bayesian bootstrapping a single database.

The reference BN (refBN) is learned from the original (uniformly weighted) database. It is the primary estimate: its posteriors are the ones returned by BootstrapMixtureInference.

The other BNs are learned from Bayesian bootstrap resamples of the same database (Dirichlet-drawn record weights simulate resampling with replacement). They are used exclusively to quantify the stability of the reference estimate: arc-confidence scores and quantile intervals on posteriors.

Parameters:
- name (str) – Name given to the reference BN. Acts as a guard: no BN added later may carry this same name.
- bn (BayesNet) – The reference BN. Every BN added later must share the same variables.