Ankur Ankan

Refer to TODOs in DAG.py:PDAG and test_PDAG.py for details.

Currently, state names are added to the parameterization object (TabularCPD or DiscreteFactor). This leads to repetition as the same state names would need to be added to all the CPDs/Factors...

Proposal for new architecture of ContinuousFactor class

34

 This structure will help us in keeping all the different distributions separate and the user will be able to use all the different distributions from ContinuousFactor only.

We are experimenting with moving the CI testing from travis and appveyor to github actions. If this works well also add actions to deploy to pypi and conda whenever a...

### Subject of the issue The current implementation of SEMs always considers the provided data to have 0 mean. Extend the implementation to handle data with non-zero means.

CPDs of models should be considered as prior in case of Bayesian Learning

6

Currently we pass in the prior probabilities when we call the learning method. Using CPDs as the prior will make the API much cleaner and easier to use.

Inference in Causal Bayesian Networks

3

### Subject of the issue Add support for causal inference in Bayesian Networks. It should be a new class accepting all the models on which causal inference can be done...

### Subject of the issue Right now the estimator methods learn all the parametrization (CPD, DiscreteFactor) for a model. This estimation should be moved to individual parameterization and the estimator...

Support for Variational inference

3

References: http://papers.nips.cc/paper/2172-vibes-a-variational-inference-engine-for-bayesian-networks.pdf http://www.jmlr.org/papers/volume6/winn05a/winn05a.pdf