Bayesian linear regression estimates a distribution over model parameters rather than only one slope. Observed data combine with a prior to produce a posterior, while predictions include uncertainty in both the parameters and observation noise. section 3.18
Predicting house prices
A simple model can relate a house’s square footage to its price, treating the slope as uncertain. The model can extend to multiple features, such as number of stories, location, and construction year, with a weight for each feature. section 3.16
Two practical ideas connect Bayesian linear regression to familiar modeling choices:
- A stronger prior concentrates the posterior toward its prior mean, which can help when data are scarce and noisy. section 3.16
- The normal–normal conjugate pair gives a normal posterior for real-valued parameters such as a regression slope. section 3.18
- Gaussian processes generalize Bayesian linear regression to infinite-degree polynomials. section 3.18
