Priors in bayesians
Noninformative priors: 1-tailed p value equals post prob for normal dist • Uniform (flat): exploratory data analysis, no prior knowledge, simple models with few parameters, improper posteriors that do not integrate to 1, no regularization • Jeffreys: some regularization to prevent overfitting with many parameters Weakly informative priors: • Normal Priors with Large Variance: L2 regularization (ridge), sensitive to the choice of variance, improper posteriors if the variance is too large, Bayesian neural network • Cauchy: default neutral(0, 0.707), pessimistic cauchy(0, 1), optimistic cauchy(0, 0.5); heavy-tailed, outliers and extreme values, somw regularization to prevent overfitting • t dist: heavy-tailed • Unit information: mean 0 sd 2, may lead to improper posteriors • Laplace: double exponential, L1 regularization (lasso), heavy-tailed • Lognormal: positive numbers, heavy-tailed • Pareto: positive numbers, heavy-tailed • Half cauchy: default for var, positive numbers, residual var,...