Jupyter Notebooks - Inference & Causality (DLMAIIAC)

This repository contains a small set of Jupyter notebooks that illustrate topics from probabilistic modelling and causal inference, including:

• Causal inference with doWhy (confounding, backdoor adjustment, interventions / the do-operator)
• Bayesian networks with pgmpy (Wet Grass / Sprinkler and the classic “Asia” network)
• Markov chains (convergence to the equilibrium / stationary distribution)
• Bayesian linear regression with PyMC3

Pinned Python dependencies are listed in requirements.txt.

Quickstart

1) Create an environment

Using venv:

python -m venv .venv
.
.venv\Scripts\activate
pip install -r requirements.txt

Using conda:

conda create -n dlmaiiuk01 python=3.8 -y
conda activate dlmaiiuk01
pip install -r requirements.txt

2) (Optional) Install Graphviz system packages

Some notebooks render graphs. For that, you may need Graphviz installed on your OS.

• Linux (Debian/Ubuntu):

sudo apt install graphviz libgraphviz-dev graphviz-dev pkg-config

On Windows, install Graphviz (e.g. from https://graphviz.org/download/) and make sure dot is available on your PATH.

3) Start JupyterLab

jupyter lab

Notebooks (what you can do with them)

• DoWhyFirstSteps.ipynb

  • Generate synthetic causal data, define a CausalModel, identify the estimand and estimate the ATE using linear regression and propensity score methods.

• doWhy_Confounder.ipynb

  • Demonstrate a classic confounding situation (“sweet spot” artifact) and show that the effect disappears after adjustment.

• doWhy_CausalDo.ipynb

  • Work with interventions using the do-operator: compare outcomes under do(X=1) vs do(X=0) and compute the ATE.

• Adjusting_for_confounder_binary_variables.ipynb

  • Binary treatment with a confounder: compare biased estimation (no confounder) vs adjusted estimation (backdoor methods) and discuss refutation.

• Adjusting_for_confounder_continuous_variables.ipynb

  • Continuous treatment and multiple outcomes: estimate effects outcome-by-outcome and run a placebo refuter to sanity-check results.

• Sprinkler.ipynb

  • Build the Wet Grass Bayesian network (C/S/R/W), define CPTs, and compute posterior probabilities using pgmpy’s variable elimination.

• AsiaBayesNet.ipynb

  • Load and query the well-known Asia Bayesian network (from a BIF file) and run inference with evidence.

• MCEquilibrium.ipynb

  • Define a Markov chain transition matrix, iterate states, and observe convergence to the equilibrium distribution.

• LinearRegressionProbabilistic.ipynb

  • Bayesian linear regression in PyMC3: specify priors, sample the posterior, inspect diagnostics, and visualize uncertainty / credible intervals.