ggplot2混淆矩阵条件填充

Question

我希望混淆矩阵具有两个不同的尺度。对于真正的积极/消极，我想绿色代表高，红色代表低。对于错误的结果，我希望颜色反转。这可能需要两个传说吗？解释起来可能很麻烦，但是我觉得倒置的颜色更直观。您的真实结果应该很高（绿色代表良好），而负面结果则应该较低（绿色代表良好）。

如果除了我所建议的以外，还有其他建议或最佳做法，我愿意接受其他选择。

library(caret)
library(tidyverse)
data("GermanCredit")
cm <- confusionMatrix(GermanCredit$Class, sample(GermanCredit$Class))

cm$table %>%
  data.frame() %>% 
  mutate(Prediction = factor(Prediction, levels = c("Good", "Bad"))) %>%
  group_by(Reference) %>% 
  mutate(total = sum(Freq)) %>% 
  ungroup() %>% 
  ggplot(aes(Reference, Prediction, fill = Freq)) +
  geom_tile() +
  geom_text(aes(label = Freq), size = 8) +
  scale_fill_gradient(low = "#ea4434", high = "#badb33") +
  scale_x_discrete(position = "top") +
  geom_tile(color = "black", fill = "black", alpha = 0)

基于一些评论，我添加了我的问题：我希望不相关的对角线为白色，而红色（低）和绿色（高）比例仅适用于真实的正/负结果。例如，下面的图部分实现了这一点，除了我将低点设置为白色...我希望低点为红色，而白色方块保持白色。

library(reprex)
library(caret)
#> Loading required package: lattice
#> Loading required package: ggplot2
library(tidyverse)
data("GermanCredit")
cm <- confusionMatrix(GermanCredit$Class, sample(GermanCredit$Class))

cm$table %>%
  data.frame() %>% 
  mutate(Prediction = factor(Prediction, levels = c("Good", "Bad"))) %>%
  group_by(Reference) %>% 
  mutate(
    total = sum(Freq),
    frac_fill = if_else(Prediction == Reference, Freq / total, 0),
    frac = Freq / total * frac_fill
  ) %>%
  ggplot(aes(Prediction, Reference, fill = frac_fill)) +
  geom_tile() +
  geom_text(aes(label = str_c(Freq, ", ", round(frac * 100), "%")), size = 8) +
  scale_fill_gradient(low = "white", high = "#badb33") +
  scale_x_discrete(position = "top") +
  geom_tile(color = "black", fill = "black", alpha = 0)

Answer 1

如果我的理解正确，您想要一个发散的比例，这需要使比例之一为负，而比例为正。如果预测为“不良”，我将变量frac_directed设为frac_fill的负数，尽管这可能与您要查找的相反。然后使用scale_fill_gradient2代替scale_fill_gradient，您可以设置一个中点（默认为0，此处合适）和中间填充值。

library(caret)
library(tidyverse)
data("GermanCredit")
cm <- confusionMatrix(GermanCredit$Class, sample(GermanCredit$Class))

cm$table %>%
  data.frame() %>% 
  mutate(Prediction = factor(Prediction, levels = c("Good", "Bad"))) %>%
  group_by(Reference) %>% 
  mutate(
    total = sum(Freq),
    frac_fill = if_else(Prediction == Reference, Freq / total, 0),
    frac = Freq / total * frac_fill
  ) %>%
  mutate(frac_directed = if_else(Prediction == "Bad", frac_fill * -1, frac_fill)) %>%
  ggplot(aes(Prediction, Reference, fill = frac_directed)) +
  geom_tile(color = "black") +
  geom_text(aes(label = str_c(Freq, ", ", round(frac * 100), "%")), size = 8) +
  scale_fill_gradient2(low = "red", mid = "white", high = "#badb33") +
  scale_x_discrete(position = "top")

由reprex package（v0.2.0）于2018-07-09创建。

Answer 2

好，这是一种尝试，主要是对非对角线图块使用反向填充值。我将轴颠倒了，以至于我通常认为是混淆矩阵。我猜这表明，该模型比Good更能预测Bad，但我认为当混淆时，它掩盖了实际混淆的内容。当这样呈现时，传说也不是最有用的信息，我什至可以不去理会。

library(caret)
#> Loading required package: lattice
#> Loading required package: ggplot2
library(tidyverse)
data("GermanCredit")
cm <- confusionMatrix(GermanCredit$Class, sample(GermanCredit$Class))

cm$table %>%
  data.frame() %>% 
  mutate(Prediction = factor(Prediction, levels = c("Good", "Bad"))) %>%
  group_by(Reference) %>% 
  mutate(
    total = sum(Freq),
    frac = Freq / total,
    frac_fill = if_else(Prediction == Reference, Freq / total, 1 - (Freq / total))
    ) %>%
  ggplot(aes(Prediction, Reference, fill = frac_fill)) +
  geom_tile() +
  geom_text(aes(label = str_c(Freq, ", ", round(frac * 100), "%")), size = 8) +
  scale_fill_gradient(low = "#ea4434", high = "#badb33") +
  scale_x_discrete(position = "top") +
  geom_tile(color = "black", fill = "black", alpha = 0)

由reprex package（v0.2.0）于2018-07-09创建。