Bayesian Logistic Regression
In addition to point estimates, we may want to measure uncertainty
- need to approximate posterior distribution (MAP: ˆw)
p(y|x,D)≈∫p(y|x,w)δ(w−ˆw)dw=p(y|x,ˆw) * comparative advantage with smaller data sets
Laplace Approximation
For a unique solution, we employ a spherical Gaussian prior
N(w|0,σ2I)
- informative prior (small σ2): better sensitivity
- vague prior (large σ2): better specificity