1. クイックスタート
base R 同梱の penguins（欠損値を自然に含む。R 4.5
以降に同梱）を使う。
library(chaidr)
data(penguins)
fit <- chaid(species ~ ., data = penguins,
control = chaid_control(min_parent = 30, min_child = 10))
print(fit)
#> CHAID decision tree (method = "chaid")
#> Response: species (categorical)
#> Valid cases: 344 (data: 344 rows)
#>
#> [1] root: Adelie (44.2%), n=344 | split: flipper_len (adj.p=2.48e-63, chi2=328.5, B=714)
#> [2] flipper_len in {<= 190}: Adelie (84.8%), n=99 | split: bill_len (adj.p=2.92e-16, chi2=78.21, B=28)
#> [3] bill_len in {<= 40.1}: Adelie (100.0%), n=70 *
#> [4] bill_len in {(40.1, 41.8]}: Adelie (92.3%), n=13 *
#> [5] bill_len in {(41.8, 47.3] | > 49.3}: Chinstrap (87.5%), n=16 *
#> [6] flipper_len in {(190, 196]}: Adelie (67.2%), n=67 | split: bill_len (adj.p=1.66e-13, chi2=58.69, B=9)
#> [7] bill_len in {<= 44.4}: Adelie (100.0%), n=43 *
#> [8] bill_len in {> 44.4}: Chinstrap (91.7%), n=24 *
#> [9] flipper_len in {(196, 202]}: Chinstrap (52.6%), n=38 | split: bill_len (adj.p=2.31e-07, chi2=30.78, B=8)
#> [10] bill_len in {<= 45.9}: Adelie (90.0%), n=20 *
#> [11] bill_len in {> 47.3}: Chinstrap (100.0%), n=18 *
#> [12] flipper_len in {(202, 214] | <NA>}: Gentoo (73.8%), n=61 | split: bill_dep (adj.p=9.69e-12, chi2=56.14, B=15)
#> [13] bill_dep in {<= 16.7}: Gentoo (100.0%), n=44 *
#> [14] bill_dep in {> 17.8 | <NA>}: Chinstrap (64.7%), n=17 *
#> [15] flipper_len in {> 214}: Gentoo (100.0%), n=79 *連続変数 flipper_len
が自動で10分位ビンに離散化され、類似ビンが統合されて
多岐分岐の木になっている。*
は末端ノード。<NA> を含むグループは、 欠損値が
Floating
カテゴリとして最も類似したビンへ自動統合されたことを示す（§4.4）。
同じ木は plot()
で可視化できる（ノード内の帯はクラス分布。詳細は §5.1）。
予測は S3 の predict で行う。
pred <- predict(fit, penguins) # クラス予測（factor）
mean(pred == penguins$species) # 訓練データ精度
#> [1] 0.9622093
round(predict(fit, head(penguins, 3), type = "prob"), 3) # クラス確率
#> Adelie Chinstrap Gentoo
#> [1,] 1 0 0
#> [2,] 1 0 0
#> [3,] 1 0 0summary()
は木に加えて設定と停止理由の内訳を表示する。
summary(fit)
#> CHAID decision tree (method = "chaid")
#> Response: species (categorical)
#> Valid cases: 344 (data: 344 rows)
#>
#> [1] root: Adelie (44.2%), n=344 | split: flipper_len (adj.p=2.48e-63, chi2=328.5, B=714)
#> [2] flipper_len in {<= 190}: Adelie (84.8%), n=99 | split: bill_len (adj.p=2.92e-16, chi2=78.21, B=28)
#> [3] bill_len in {<= 40.1}: Adelie (100.0%), n=70 *
#> [4] bill_len in {(40.1, 41.8]}: Adelie (92.3%), n=13 *
#> [5] bill_len in {(41.8, 47.3] | > 49.3}: Chinstrap (87.5%), n=16 *
#> [6] flipper_len in {(190, 196]}: Adelie (67.2%), n=67 | split: bill_len (adj.p=1.66e-13, chi2=58.69, B=9)
#> [7] bill_len in {<= 44.4}: Adelie (100.0%), n=43 *
#> [8] bill_len in {> 44.4}: Chinstrap (91.7%), n=24 *
#> [9] flipper_len in {(196, 202]}: Chinstrap (52.6%), n=38 | split: bill_len (adj.p=2.31e-07, chi2=30.78, B=8)
#> [10] bill_len in {<= 45.9}: Adelie (90.0%), n=20 *
#> [11] bill_len in {> 47.3}: Chinstrap (100.0%), n=18 *
#> [12] flipper_len in {(202, 214] | <NA>}: Gentoo (73.8%), n=61 | split: bill_dep (adj.p=9.69e-12, chi2=56.14, B=15)
#> [13] bill_dep in {<= 16.7}: Gentoo (100.0%), n=44 *
#> [14] bill_dep in {> 17.8 | <NA>}: Chinstrap (64.7%), n=17 *
#> [15] flipper_len in {> 214}: Gentoo (100.0%), n=79 *
#>
#> Settings: alpha_merge=0.05, alpha_split=0.05, bonferroni=TRUE, adjust_across=none, max_depth=3, min_parent=30, min_child=10, n_bins=10
#>
#> Risk estimate (training data): 0.0378 (misclassification rate)
#>
#> Terminal nodes: 10
#> Stopping reasons:
#> min_parent: 5
#> pure: 5
