Data disimulasikan dengan hubungan non-linier (kuadratik) antara x dan y.
set.seed(123)
n <- 100
x <- seq(0, 10, length.out = n)
y <- 5 + 2*x - 0.3*x^2 + rnorm(n, mean = 0, sd = 2)
data <- data.frame(x = x, y = y)
# Menampilkan 10 baris pertama dari 100 pengamatan
head(data, 10)
Tiga model dicocokkan sebagai pembanding: linier, polinomial derajat 2, dan derajat 3.
model_linier <- lm(y ~ x, data = data)
model_poli2 <- lm(y ~ poly(x, 2, raw = TRUE), data = data)
model_poli3 <- lm(y ~ poly(x, 3, raw = TRUE), data = data)
summary(model_linier)
##
## Call:
## lm(formula = y ~ x, data = data)
##
## Residuals:
## Min 1Q Median 3Q Max
## -7.4317 -1.4051 -0.0466 2.2166 6.6109
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 9.88171 0.55369 17.847 <2e-16 ***
## x -0.95028 0.09566 -9.934 <2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 2.789 on 98 degrees of freedom
## Multiple R-squared: 0.5017, Adjusted R-squared: 0.4967
## F-statistic: 98.68 on 1 and 98 DF, p-value: < 2.2e-16
summary(model_poli2)
##
## Call:
## lm(formula = y ~ poly(x, 2, raw = TRUE), data = data)
##
## Residuals:
## Min 1Q Median 3Q Max
## -4.8136 -1.1977 -0.0533 1.3549 4.3891
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 5.33982 0.53777 9.929 <2e-16 ***
## poly(x, 2, raw = TRUE)1 1.80266 0.24855 7.253 1e-10 ***
## poly(x, 2, raw = TRUE)2 -0.27529 0.02405 -11.447 <2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 1.829 on 97 degrees of freedom
## Multiple R-squared: 0.788, Adjusted R-squared: 0.7837
## F-statistic: 180.3 on 2 and 97 DF, p-value: < 2.2e-16
summary(model_poli3)
##
## Call:
## lm(formula = y ~ poly(x, 3, raw = TRUE), data = data)
##
## Residuals:
## Min 1Q Median 3Q Max
## -4.7866 -1.1996 -0.0497 1.3386 4.3777
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 5.282801 0.708433 7.457 3.93e-11 ***
## poly(x, 3, raw = TRUE)1 1.872854 0.616613 3.037 0.00307 **
## poly(x, 3, raw = TRUE)2 -0.292931 0.143693 -2.039 0.04424 *
## poly(x, 3, raw = TRUE)3 0.001176 0.009443 0.125 0.90117
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 1.838 on 96 degrees of freedom
## Multiple R-squared: 0.7881, Adjusted R-squared: 0.7815
## F-statistic: 119 on 3 and 96 DF, p-value: < 2.2e-16
anova(model_linier, model_poli2, model_poli3)
AIC(model_linier, model_poli2, model_poli3)
data$pred_linier <- predict(model_linier)
data$pred_poli2 <- predict(model_poli2)
data$pred_poli3 <- predict(model_poli3)
ggplot(data, aes(x = x, y = y)) +
geom_point(alpha = 0.5) +
geom_line(aes(y = pred_linier, color = "Linier"), linewidth = 1) +
geom_line(aes(y = pred_poli2, color = "Polinomial derajat 2"), linewidth = 1) +
geom_line(aes(y = pred_poli3, color = "Polinomial derajat 3"), linewidth = 1) +
labs(title = "Perbandingan Regresi Linier vs Polinomial",
x = "X", y = "Y", color = "Model") +
theme_minimal()
RMSE di bagian ini dihitung pada data yang sama dengan data pelatihan (in-sample), sehingga nilainya tidak pernah naik ketika derajat ditambah. Karena itu RMSE ini belum cocok untuk memilih derajat, dan pemilihannya dilakukan lewat cross-validation pada bagian berikutnya.
rmse <- function(actual, predicted) sqrt(mean((actual - predicted)^2))
data.frame(
Model = c("Linier", "Polinomial derajat 2", "Polinomial derajat 3"),
RMSE = c(rmse(data$y, data$pred_linier),
rmse(data$y, data$pred_poli2),
rmse(data$y, data$pred_poli3))
)
Skema 10-fold cross-validation diulang 5 kali untuk menguji derajat polinomial 1 sampai 6.
kontrol <- trainControl(
method = "repeatedcv",
number = 10,
repeats = 5
)
derajat_max <- 6
hasil_cv <- data.frame(
derajat = integer(),
RMSE = numeric()
)
for (d in 1:derajat_max) {
data_cv <- data[, c("x", "y")]
# Membuat variabel polynomial
if (d == 1) {
data_cv$x_poly <- data_cv$x
} else {
poly_x <- poly(data_cv$x, degree = d, raw = TRUE)
data_cv$x_poly <- poly_x[, 1]
for (j in 2:d) {
data_cv[[paste0("x_poly", j)]] <- poly_x[, j]
}
}
# Formula sesuai derajat
if (d == 1) {
formula_model <- y ~ x_poly
} else {
variabel <- paste0("x_poly", 2:d)
formula_model <- as.formula(
paste("y ~ x_poly +", paste(variabel, collapse = " + "))
)
}
# Seed yang sama untuk setiap derajat agar pembagian fold-nya identik
set.seed(42)
model_cv <- train(
formula_model,
data = data_cv,
method = "lm",
trControl = kontrol
)
hasil_cv <- rbind(
hasil_cv,
data.frame(
derajat = d,
RMSE = min(model_cv$results$RMSE)
)
)
}
hasil_cv
derajat_optimal <- hasil_cv$derajat[which.min(hasil_cv$RMSE)]
cat("Derajat optimal berdasarkan CV:", derajat_optimal, "\n")
## Derajat optimal berdasarkan CV: 2
ggplot(hasil_cv, aes(x = derajat, y = RMSE)) +
geom_line(
color = "steelblue",
linewidth = 1
) +
geom_point(
size = 3,
color = "steelblue"
) +
geom_vline(
xintercept = derajat_optimal,
linetype = "dashed",
color = "firebrick"
) +
scale_x_continuous(
breaks = 1:derajat_max
) +
labs(
title = "Pemilihan Derajat Optimal via 10-Fold Cross-Validation",
x = "Derajat Polinomial",
y = "RMSE Rata-rata (Validasi Silang)"
) +
theme_minimal()
model_final <- lm(y ~ poly(x, derajat_optimal, raw = TRUE), data = data)
summary(model_final)
##
## Call:
## lm(formula = y ~ poly(x, derajat_optimal, raw = TRUE), data = data)
##
## Residuals:
## Min 1Q Median 3Q Max
## -4.8136 -1.1977 -0.0533 1.3549 4.3891
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 5.33982 0.53777 9.929 <2e-16 ***
## poly(x, derajat_optimal, raw = TRUE)1 1.80266 0.24855 7.253 1e-10 ***
## poly(x, derajat_optimal, raw = TRUE)2 -0.27529 0.02405 -11.447 <2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 1.829 on 97 degrees of freedom
## Multiple R-squared: 0.788, Adjusted R-squared: 0.7837
## F-statistic: 180.3 on 2 and 97 DF, p-value: < 2.2e-16
model_linier <- lm(y ~ x, data = data)
model_poli2 <- lm(y ~ x + I(x^2), data = data)
model_poli3 <- lm(y ~ x + I(x^2) + I(x^3), data = data)
data$y_topi_linier <- predict(model_linier)
data$y_topi_poli2 <- predict(model_poli2)
data$y_topi_poli3 <- predict(model_poli3)
data
poly(x, derajat, raw = TRUE) digunakan agar
koefisien dapat diinterpretasi langsung sebagai \(\beta_1 x + \beta_2 x^2 + \dots\)
Untuk derajat tinggi (\(\geq
4\)), sebaiknya x distandardisasi terlebih dahulu
(scale(x)) agar model lebih stabil secara numerik.
Pertimbangkan aturan one-standard-error rule: pilih derajat terkecil dengan RMSE \(\leq\) (RMSE minimum + 1 standard error) untuk model yang lebih sederhana namun performanya setara.