In haziqj/iprobit: Binary and Multinomial Probit Regression with I-priors

knitr::opts_chunk$set(echo = TRUE, collapse = TRUE)
library(iprobit)

This is an R package which extends I-prior regression to unordered categorical responses via a probit link function. This allows the user to fit models for classification or inference using fitted probabilities. Estimation is performed using a variational EM algorithm. Visit http://phd.haziqj.ml for details.

Binary classification (toy example)

Model fitting

dat <- gen_spiral(n = 300, seed = 123)  # generate binary toy example data set
mod <- iprobit(y ~ X1 + X2, dat, one.lam = TRUE, kernel = "fbm")

Model summary

summary(mod)

Boundary plot for two-dimensional covariates

iplot_predict(mod)

Multiclass classification (toy example)

Model fit report and parameter estimates

dat <- gen_mixture(n = 400, m = 4, sd = 1.5, seed = 123)  # generate 4-class 
                                                          # toy example data set
(mod <- iprobit(y ~ X1 + X2, dat, train.samp = sample(1:400, size = 392),
                control = list(maxit = 10)))  # set aside 8 points for testing

Boundary plot for two-dimensional covariates

iplot_predict(mod, dec.bound = TRUE, plot.test = TRUE, grid.len = 50)

Obtain out-of-sample test error rates, predicted classes and probabilities

predict(mod)

Fisher's Iris data set

Model fitting (common RKHS scale across classes for each covariate)

set.seed(123)
mod <- iprobit(Species ~ ., iris, kernel = "fbm", one.lam = TRUE,
               common.RKHS.scale = TRUE, common.intercept = FALSE,
               control = list(alpha0 = 1, theta0 = 1, 
                              stop.crit = 1e-1))

summary(mod)

Obtain training error rates, predicted classes and probabilities with posterior quantiles

fitted(mod, quantiles = TRUE)

Monitor convergence

iplot_lb(mod)

Plot of training error over time

iplot_error(mod)

Plot of fitted probabilities

iplot_fitted(mod)

---

Copyright (C) 2017 Haziq Jamil.

haziqj/iprobit documentation built on May 24, 2019, 4:05 a.m.