Bayesian Linear Regression

Linear regression is based on the assumption that the underlying data is normally distributed and that all relevant predictor variables have a linear relationship with the outcome.  But In the real world, this is not always possible, it will follows these assumptions, Bayesian regression could be the better choice.

Why Bayesian Regression Can Be a Better Choice?

Bayesian regression employs prior belief or knowledge about the data to "learn" more about it and create more accurate predictions. It also takes into account the data's uncertainty and leverages prior knowledge to provide more precise estimates of the data. As a result, it is an ideal choice when the data is complex or ambiguous.

Bayesian regression leverages Bayes' theorem to estimate the parameters of a linear model, incorporating both observed data and prior beliefs about the parameters. Unlike ordinary least squares (OLS) regression, which provides point estimates, Bayesian regression produces probability distributions over possible parameter values, offering a measure of uncertainty in predictions.

Core Concepts in Bayesian Regression

The important concepts in Bayesian Regression are as follows:

Bayes’ Theorem

Bayes’ theorem describes how prior knowledge is updated with new data:

where:

• P(A|B) is the posterior probability after observing data.
• P(B|A) is the likelihood of the data given the parameters.
• P(A) is the prior probability.
• P(B) is the marginal probability of the observed data.

Likelihood Function

The likelihood function represents the probability of the observed data given certain parameter values. Assuming normal errors, the relationship between independent variables X and target variable Y is:

where follows a normal distribution variance  .

Prior and Posterior Distributions

• Prior P( w ∣ α): Represents prior knowledge about the parameters before observing data.
• Posterior P( w ∣ X ,α ,β⁻¹ ): Updated beliefs about the parameters after incorporating observed data, derived using Bayes’ theorem.

Need for Bayesian Regression

Bayesian regression offers several advantages over traditional regression techniques:

1. Handles Limited Data: Incorporates prior knowledge, making it effective when data is scarce.
2. Uncertainty Estimation: Provides a probability distribution for parameters rather than a single estimate, allowing for credible intervals.
3. Flexibility: Accommodates complex relationships and non-linearities by integrating various prior distributions.
4. Model Selection: Enables comparison of different models using posterior probabilities.
5. Robust to Outliers: Reduces the influence of extreme values, making it more stable than classical regression methods.

Bayesian Regression Formulation

For a dataset with n samples, the linear relationship is:

where w are regression coefficients and .

Assumptions:

1. The error terms are independent and identically distributed (i.i.d.).
2. The target variable Y follows a normal distribution with mean and variance σ², i.e.,

Conditional Probability Density Function (PDF)

The probability density function of Y given X is:

For N observations:

which simplifies to:

Taking the logarithm of the likelihood function:

Precision Term

We define precision β as:

Substituting into the likelihood function:

The negative log-likelihood is:

Maximum Posterior Estimation

Taking the logarithm of the posterior:

Substituting the expressions:

Minimizing this expression gives the maximum posterior estimate, which is equivalent to ridge regression.

Bayesian regression provides a probabilistic framework for linear regression by incorporating prior knowledge. Instead of estimating a single set of parameters, we obtain a distribution over possible parameters, which enhances robustness in situations with limited data or multicollinearity.

Key Differences: Traditional vs. Bayesian Regression

When to Use Bayesian Regression?

• Small sample sizes: When data is scarce, Bayesian inference can improve predictions.
• Strong prior knowledge: When domain expertise is available, incorporating priors enhances model reliability.
• Handling uncertainty: If quantifying uncertainty in predictions is essential.

Implementation of Bayesian Regression Using Python

Method 1: Bayesian Linear Regression using Stochastic Variational Inference (SVI) in Pyro.

It utilizes Stochastic Variational Inference (SVI) to approximate the posterior distribution of parameters (slope, intercept, and noise variance) in a Bayesian linear regression model. The Adam optimizer is used to minimize the Evidence Lower Bound (ELBO), making the inference computationally efficient.

Step 1: Import Required Libraries

First, we import the necessary Python libraries for performing Bayesian regression using torch, pyro, SVI, Trace_ELBO, predictive, Adam, and matplotlib and seaborn.

Step 2: Generate Sample Data

We create synthetic data for linear regression:

• Y = intercept + slope × X + noise
• The noise follows a normal distribution to simulate real-world uncertainty.

Step 3: Define the Bayesian Regression Model

• Priors: Assign normal distributions to the slope and intercept.
• Likelihood: The observations Y follow a normal distribution centered around μ = intercept + slope × X .

Step 4: Define the Variational Guide

This function approximates the posterior distribution of the parameters:

• Usespyro.param to learn mean (loc) and standard deviation (scale) for each parameter.
• Samples are drawn from these learned distributions.

Step 5: Train the Model using SVI

• Adam optimizer is used for parameter updates.
• SVI minimizes the ELBO (Evidence Lower Bound) to approximate the posterior.

Output

Iteration 100/1000 - Loss: 857.5039891600609
Iteration 200/1000 - Loss: 76392.14724761248
Iteration 300/1000 - Loss: 4466.2376717329025
Iteration 400/1000 - Loss: 70616.07956075668
Iteration 500/1000 - Loss: 7564.8086141347885
Iteration 600/1000 - Loss: 86843.96660631895
Iteration 700/1000 - Loss: 155.43085688352585
Iteration 800/1000 - Loss: 248.03456103801727
Iteration 900/1000 - Loss: 353587.08260041475
Iteration 1000/1000 - Loss: 253.0774005651474

Step 6: Obtain Posterior Samples

• Predictive function samples from the posterior using the trained guide.
• We extract samples for slope, intercept, and sigma.

Output

Estimated Parameters:
Estimated Slope: 1.0719
Estimated Intercept: 1.1454
Estimated Sigma: 2.2641

Step 7: Compute and Display Results

We plot the distributions of the inferred parameters: slope, intercept, and sigma using seaborn

Output

Method: 2 Bayesian Linear Regression using PyMC3

In this implementation, we utilize Bayesian Linear Regression with Markov Chain Monte Carlo (MCMC) sampling using PyMC3, allowing for a probabilistic interpretation of regression parameters and their uncertainties.

1. Import Necessary Libraries

Here, we import the required libraries for the task. These libraries include os, pytensor, pymc, numpy, and matplotlib.

2. Clear PyTensor Cache

PyMC uses PyTensor (formerly Theano) as the backend for running computations. We clear the cache to avoid any potential issues with stale compiled code

3. Set Random Seed and Generate Synthetic Data

We combine setting the random seed and generating synthetic data in this step. The random seed ensures reproducibility, and the synthetic data is generated for the linear regression model.

4. Define the Bayesian Model

Now, we define the Bayesian model using PyMC. Here, we specify the priors for the model parameters (slope, intercept, and sigma), and the likelihood function for the observed data.

5. Sample from the Posterior

After defining the model, we sample from the posterior using MCMC (Markov Chain Monte Carlo). The pm.sample() function draws samples from the posterior distributions of the model parameters.

• We set draws=2000 for the number of samples, tune=1000 for tuning steps, and cores=1 to use a single core for the sampling process.

6. Plot the Posterior Distributions

Finally, we plot the posterior distributions of the parameters (slope, intercept, and sigma) to visualize the uncertainty in their estimates. pm.plot_posterior()plots the distributions, showing the most likely values for each parameter.

Output

Advantages of Bayesian Regression

• Effective for small datasets: Works well when data is limited.
• Handles uncertainty: Provides probability distributions instead of point estimates.
• Flexible modeling: Can handle complex relationships and non-linearity.
• Robust against outliers: Unlike OLS, Bayesian regression reduces the impact of extreme values.
• Facilitates model selection: Computes posterior probabilities for different models.

Limitations of Bayesian Regression

• Computationally expensive: Requires advanced sampling techniques.
• Requires specifying priors: Poorly chosen priors can affect results.
• Not always necessary: For large datasets, traditional regression often performs adequately.