混淆矩阵没有显示实际值的正确计数。多项回归,因素

时间:2018-05-23 22:19:38

标签: r confusion-matrix multinomial

我有两个向量,实际值和预测值。两者都是8级的因子类型。第8级实际上只有55个观测值,预测值为0。然而,当我制作混淆矩阵时,8级观察结果消失或以某种方式移动。不应该将实际总和的列与实际计数相对应吗?

我制作了混淆矩阵两种不同的方法来仔细检查。我还尝试明确地使两个向量中的因子水平相同。到目前为止没有运气。

library(nnet); library(caret)

sc <- read.csv("https://archive.ics.uci.edu/ml/machine-learning-databases/00272/SkillCraft1_Dataset.csv")

# First column is ID
sc$LeagueIndex <- as.factor(sc$LeagueIndex)
sc <- sc[, -1]

# Set missing values to NA
which_qm <- sc[, c(2,3,4)] == '?'
sc[, c(2,3,4)][which_qm] <- NA
sc[, c(2,3,4)] <- apply(sc[, c(2,3,4)], 2, as.numeric)

# Set impossible values to NA
sc$TotalHours[sc$Age < sc$TotalHours/8760] <- NA
sc$HoursPerWeek[sc$HoursPerWeek >= 168] <- NA

# Fit model and store predictions
sc_mod1 <- multinom(LeagueIndex ~ ., sc)
sc_fitted1 <- predict(sc_mod1, sc)

# sc_fitted1 is missing factor level 8
confusionMatrix(data = sc_fitted1, reference = sc$LeagueIndex)
table(predicted = sc_fitted1, actual = sc$LeagueIndex)

# sc_fitted1 has factor level 8
levels(sc_fitted1) <- levels(sc$LeagueIndex)
confusionMatrix(data = sc_fitted1, reference = sc$LeagueIndex)
table(predicted = sc_fitted1, actual = sc$LeagueIndex)

# What's the problem?
table(sc$LeagueIndex)
length(sc$LeagueIndex)

table(sc_fitted1)
length(sc_fitted1)

1 个答案:

答案 0 :(得分:1)

它与您生成的NA值有关,它们都是目标变量的8级。如果你想要考虑第8级,你可能必须找到另一种编码这些NA的方法。

试试这个作为反例:

library(nnet); library(caret)

sc <- read.csv("https://archive.ics.uci.edu/ml/machine-learning-databases/00272/SkillCraft1_Dataset.csv")

sc$LeagueIndex <- as.factor(sc$LeagueIndex)
sc <- sc[, -1]

which_qm <- sc[, c(2,3,4)] == '?'
sc[, c(2,3,4)][which_qm] <- 20   # this is just a random numeric value (not the best one to use!)
sc[, c(2,3,4)] <- apply(sc[, c(2,3,4)], 2, as.numeric)

sc_mod1 <- multinom(LeagueIndex ~ ., sc)
sc_fitted1 <- predict(sc_mod1, sc)

confusionMatrix(data = sc_fitted1, reference = sc$LeagueIndex)
table(predicted = sc_fitted1, actual = sc$LeagueIndex)

它会给你这样的东西:

         actual
predicted   1   2   3   4   5   6   7   8
        1  52  30   9   2   0   0   0   0
        2  61 123  78  58   4   1   0   0
        3  30  77 142  79  23   4   0   0
        4  21 104 248 410 252  45   0   0
        5   2  11  60 217 343 230   1   0
        6   1   2  16  45 184 333  32   2
        7   0   0   0   0   0   5   2   0
        8   0   0   0   0   0   3   0  53