Skip to content

Latest commit

 

History

History
67 lines (52 loc) · 2.64 KB

svm_gridsearch.md

File metadata and controls

67 lines (52 loc) · 2.64 KB
Raw

Grid search results for SVM

Data

library(tidyverse)
library(here)
d = read_csv(here("data/intermediate/svm_grid.csv"))
d = d %>% group_by(experiment, kernel, C, gamma) %>% summarize(m=mean(accuracy), sd=sd(accuracy))

Best model

d %>% arrange(-m)
## # A tibble: 585 x 6
## # Groups:   experiment, kernel, C [585]
##    experiment kernel     C gamma                   m     sd
##         <dbl> <chr>  <dbl> <chr>               <dbl>  <dbl>
##  1        296 rbf     1    1.0                 0.579 0.0104
##  2        272 rbf     1    scale               0.579 0.0108
##  3        338 rbf     3.16 0.31622776601683794 0.574 0.0127
##  4        377 rbf    10    0.03162277660168379 0.572 0.0115
##  5        312 linear  1    1000.0              0.572 0.0121
##  6        306 linear  1    100.0               0.572 0.0113
##  7        291 linear  1    0.31622776601683794 0.572 0.0120
##  8        419 rbf    31.6  0.01                0.571 0.0114
##  9        270 linear  1    scale               0.571 0.0131
## 10        273 linear  1    auto                0.571 0.0109
## # … with 575 more rows

Parameters

d = d %>% mutate(gamma2=case_when(gamma %in% c("scale", "auto") ~ gamma, T~as.character(log10(as.numeric(gamma)))))

d %>% group_by(kernel, gamma) %>% arrange(-m) %>% filter(row_number() == 1) %>% 
  ggplot(aes(x=fct_reorder(gamma2, as.numeric(gamma2)), y=m, group=kernel, color=kernel)) + geom_line() +
  ggthemes::theme_clean() + xlab("log10(gamma)") + ylab("Accuracy") + ggtitle("Effect of kernel and gamma", "Setting C to maximize accuracy per experiment")

d %>% group_by(kernel, C) %>% arrange(-m) %>% filter(row_number() == 1) %>% 
  ggplot(aes(x=log10(C), y=m, group=kernel, color=kernel)) + geom_line() +
  ggthemes::theme_clean() + xlab("log10(C)") + ylab("Accuracy") + ggtitle("Effect of kernel and C", "Setting gamma to maximize accuracy per experiment")

d %>% filter(C==1) %>% ggplot(aes(x=fct_reorder(gamma2, as.numeric(gamma2)), y=m, group=kernel, color=kernel)) + geom_line() +
  ggthemes::theme_clean() + xlab("log10(gamma)") + ylab("Accuracy") + ggtitle("Effect of kernel and gamma", "Fixing C=1 (global maximum)")

d %>% filter(gamma=="1.0") %>% ggplot(aes(x=log10(C), y=m, group=kernel, color=kernel)) + geom_line() +
  ggthemes::theme_clean() + xlab("log10(C)") + ylab("Accuracy") + ggtitle("Effect of kernel and C", "Fixing gamma=1 (global maximum)")