Notebook Bab 3 - Representasi Fitur Numerik

Open In Colab

Notebook Bab 3 ini punya dua bagian. Bagian Demo tinggal Anda jalankan lalu amati keluarannya; bagian Mini Project berisi soal dan data yang Anda kerjakan sendiri.

Skala fitur numerik memengaruhi model berbasis jarak (k-NN, SVM), tetapi tidak model berbasis pohon. Di sini kita mengukur efeknya pada data yang sama.

Persiapan

import os
os.environ.setdefault('LOKY_MAX_CPU_COUNT', '4')

import json
from pathlib import Path

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from sklearn.datasets import make_classification
from sklearn.ensemble import HistGradientBoostingRegressor
from sklearn.metrics import mean_squared_error
from sklearn.model_selection import train_test_split
from sklearn.neighbors import KNeighborsClassifier, KNeighborsRegressor
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import MinMaxScaler, PowerTransformer, QuantileTransformer, RobustScaler, StandardScaler

RANDOM_STATE = 42
print('Setup selesai.')


from pathlib import Path
import json
import urllib.request
import urllib.parse

DATA_BASE_URL = 'https://raw.githubusercontent.com/muhammad-zainal-muttaqin/NulisBuku/main/website/notebooks/data/section1'


def section_data_dir(name):
    """Folder data Bagian 1: pakai salinan lokal bila ada; jika tidak (mis. di
    Google Colab), unduh berkas dari repo GitHub sesuai manifest."""
    for base in (Path('data/section1'), Path('../data/section1')):
        if (base / name).exists():
            return base / name
    cache = Path('_nb_data') / name
    if not (cache / 'manifest.json').exists():
        cache.mkdir(parents=True, exist_ok=True)
        base_url = DATA_BASE_URL + '/' + name
        manifest = json.loads(urllib.request.urlopen(base_url + '/manifest.json').read().decode('utf-8'))
        for rel in manifest:
            dest = cache / rel
            dest.parent.mkdir(parents=True, exist_ok=True)
            if not dest.exists():
                url = base_url + '/' + '/'.join(urllib.parse.quote(seg) for seg in rel.split('/'))
                urllib.request.urlretrieve(url, dest)
        (cache / 'manifest.json').write_text(json.dumps(manifest), encoding='utf-8')
    return cache

Setup selesai.

Section 1 - Demo: Transformasi Numerik pada Online News Popularity

Data nyata: fitur numerik artikel berita daring

DATA_DIR = section_data_dir('ch03_online_news_popularity')
df = pd.read_parquet(DATA_DIR / 'online_news_popularity.parquet')
stats = json.loads((DATA_DIR / 'verified_stats.json').read_text(encoding='utf-8'))

feature_cols = [c for c in df.columns if c not in ['url', 'timedelta', 'shares']]
shape_feature = 'kw_avg_avg'
boundary_features = ['num_hrefs', 'kw_avg_avg']

print('Snapshot:', DATA_DIR / 'online_news_popularity.parquet')
print(f"Baris: {stats['rows']:,} | Fitur prediktif numerik: {stats['predictive_feature_count']} | Missing cells: {stats['missing_cells']}")
print(f"Target shares: median={stats['shares_median']:,.0f}, max={stats['shares_max']:,}, skew={stats['shares_skew']:.3f}, IQR outliers={stats['shares_iqr_outliers']:,}")
print('Catatan:', stats['source_page_rows_note'])

range_table = pd.DataFrame(stats['feature_range_examples']).T[['min', 'max', 'range']]
print('\nContoh rentang fitur terbesar:')
print(range_table.head(8).round(3).to_string())

print(f"\nFitur bentuk distribusi `{shape_feature}`: min={df[shape_feature].min():.3f}, median={df[shape_feature].median():.3f}, max={df[shape_feature].max():.3f}, skew={df[shape_feature].skew():.3f}")

Snapshot: data\section1\ch03_online_news_popularity\online_news_popularity.parquet
Baris: 39,644 | Fitur prediktif numerik: 58 | Missing cells: 0
Target shares: median=1,400, max=843,300, skew=33.964, IQR outliers=4,541
Catatan: UCI web metadata currently displays 39,797 instances, while the downloaded CSV contains 39,644 data rows.

Contoh rentang fitur terbesar:
                            min       max     range
kw_avg_max                  0.0  843300.0  843300.0
kw_min_max                  0.0  843300.0  843300.0
kw_max_max                  0.0  843300.0  843300.0
self_reference_avg_sharess  0.0  843300.0  843300.0
self_reference_max_shares   0.0  843300.0  843300.0
self_reference_min_shares   0.0  843300.0  843300.0
kw_max_avg                  0.0  298400.0  298400.0
kw_max_min                  0.0  298400.0  298400.0

Fitur bentuk distribusi `kw_avg_avg`: min=0.000, median=2870.075, max=43567.660, skew=5.760

Distribusi, transformasi bentuk, dan fixed ruler

train_df, test_df = train_test_split(df, test_size=0.30, random_state=RANDOM_STATE)

# Gambar 3.3: min-max vs robust scaling pada fitur yang berekor kanan.
values = train_df[[shape_feature]].to_numpy()
minmax_values = MinMaxScaler().fit_transform(values).ravel()
robust_values = RobustScaler().fit_transform(values).ravel()
rng = np.random.default_rng(RANDOM_STATE)
sample_idx = rng.choice(len(values), size=1600, replace=False)

fig, axes = plt.subplots(3, 1, figsize=(8, 4.8), sharey=True)
for ax, arr, title in [
    (axes[0], values.ravel(), f'Mentah: {shape_feature}'),
    (axes[1], minmax_values, 'Min-max scaling'),
    (axes[2], robust_values, 'Robust scaling (median/IQR)'),
]:
    y_jitter = rng.normal(0, 0.025, size=len(sample_idx))
    ax.scatter(arr[sample_idx], y_jitter, s=10, alpha=0.35, color='0.25')
    ax.axvline(np.median(arr), color='tab:blue', lw=1.2, label='median')
    ax.set_title(title)
    ax.set_yticks([])
axes[0].legend(loc='upper right')
fig.suptitle('Gambar 3.3 - Nilai ekstrem tetap ada, tetapi pusat/skala berubah')
plt.tight_layout()

# Gambar 3.4: transformasi bentuk distribusi pada fitur prediktif real.
feature_train = train_df[[shape_feature]]
shape_views = {
    'Mentah': feature_train[shape_feature].to_numpy(),
    'log1p': np.log1p(feature_train[shape_feature].to_numpy()),
    'Yeo-Johnson': PowerTransformer(method='yeo-johnson').fit_transform(feature_train).ravel(),
    'Quantile -> normal': QuantileTransformer(
        n_quantiles=min(1000, len(feature_train)),
        output_distribution='normal',
        random_state=RANDOM_STATE,
    ).fit_transform(feature_train).ravel(),
}
fig, axes = plt.subplots(2, 2, figsize=(8, 5.2))
for ax, (name, arr) in zip(axes.ravel(), shape_views.items()):
    ax.hist(arr, bins=40, color='0.35', edgecolor='white')
    ax.set_title(name)
    ax.set_ylabel('Frekuensi')
fig.suptitle(f'Gambar 3.4 - Transformasi bentuk distribusi {shape_feature}')
plt.tight_layout()

# Fixed ruler: model sama, representasi berubah.
X_train, X_test = train_df[feature_cols], test_df[feature_cols]
y_train, y_test = np.log1p(train_df['shares']), np.log1p(test_df['shares'])

def rmse(model):
    model.fit(X_train, y_train)
    pred = model.predict(X_test)
    return np.sqrt(mean_squared_error(y_test, pred))

results = pd.DataFrame([
    {'model': 'k-NN', 'representasi': 'raw', 'RMSE_log1p_shares': rmse(KNeighborsRegressor(n_neighbors=25))},
    {'model': 'k-NN', 'representasi': 'StandardScaler', 'RMSE_log1p_shares': rmse(Pipeline([('scaler', StandardScaler()), ('model', KNeighborsRegressor(n_neighbors=25))]))},
    {'model': 'HistGradientBoosting', 'representasi': 'raw', 'RMSE_log1p_shares': rmse(HistGradientBoostingRegressor(max_iter=120, learning_rate=0.06, random_state=RANDOM_STATE))},
    {'model': 'HistGradientBoosting', 'representasi': 'StandardScaler', 'RMSE_log1p_shares': rmse(Pipeline([('scaler', StandardScaler()), ('model', HistGradientBoostingRegressor(max_iter=120, learning_rate=0.06, random_state=RANDOM_STATE))]))},
])
print(results.round(4).to_string(index=False))

               model   representasi  RMSE_log1p_shares
                k-NN            raw             0.9083
                k-NN StandardScaler             0.8894
HistGradientBoosting            raw             0.8494
HistGradientBoosting StandardScaler             0.8494

# Gambar 3.6: batas keputusan k-NN pada dua fitur real dari Online News Popularity.
y_high = (df['shares'] >= df['shares'].quantile(0.75)).astype(int)
X2 = df[boundary_features]
X2_train, X2_test, y2_train, y2_test = train_test_split(
    X2, y_high, test_size=0.30, stratify=y_high, random_state=RANDOM_STATE
)

raw_clf = KNeighborsClassifier(n_neighbors=35).fit(X2_train, y2_train)
scaled_clf = Pipeline([
    ('scaler', StandardScaler()),
    ('model', KNeighborsClassifier(n_neighbors=35)),
]).fit(X2_train, y2_train)

# Plot the central feature range so the boundary is readable despite heavy tails.
x_min, x_max = X2['num_hrefs'].quantile([0.01, 0.99])
y_min, y_max = X2['kw_avg_avg'].quantile([0.01, 0.99])
xx, yy = np.meshgrid(np.linspace(x_min, x_max, 180), np.linspace(y_min, y_max, 180))
grid = pd.DataFrame({'num_hrefs': xx.ravel(), 'kw_avg_avg': yy.ravel()})

plot_rows = X2_test[
    X2_test['num_hrefs'].between(x_min, x_max) & X2_test['kw_avg_avg'].between(y_min, y_max)
].sample(900, random_state=RANDOM_STATE)

fig, axes = plt.subplots(1, 2, figsize=(9, 4), sharex=True, sharey=True)
for ax, model, title in [
    (axes[0], raw_clf, 'Raw feature scale'),
    (axes[1], scaled_clf, 'With StandardScaler'),
]:
    zz = model.predict_proba(grid)[:, 1].reshape(xx.shape)
    ax.contourf(xx, yy, zz, levels=np.linspace(0, 1, 11), cmap='Blues', alpha=0.75)
    ax.scatter(
        plot_rows['num_hrefs'], plot_rows['kw_avg_avg'],
        c=y2_test.loc[plot_rows.index], cmap='coolwarm', s=10, alpha=0.55, edgecolor='none'
    )
    ax.set_title(title)
    ax.set_xlabel('num_hrefs')
axes[0].set_ylabel('kw_avg_avg')
fig.suptitle('Gambar 3.6 - Batas keputusan k-NN berubah setelah standardisasi')
plt.tight_layout()

print(f"Akurasi ilustratif raw={raw_clf.score(X2_test, y2_test):.3f}; scaled={scaled_clf.score(X2_test, y2_test):.3f}")

Akurasi ilustratif raw=0.743; scaled=0.746

🔎 Amati. Online News memuat fitur numerik dengan rentang yang sangat berbeda: sebagian berupa hitungan, sebagian berupa statistik popularitas historis, dan sebagian berupa rasio 0..1. Fitur kw_avg_avg berekor kanan, sehingga log, Yeo-Johnson, dan quantile transform mengubah bentuk distribusinya, bukan sekadar satuan sumbu. Pada fixed ruler, standardisasi menurunkan RMSE k-NN pada target log1p(shares), sedangkan model pohon boosting jauh lebih stabil terhadap scaling linear.

Section 2 - Mini Project

Soal

Anda diberi data numerik dengan distribusi menceng dan skala antar-kolom yang timpang. Bangun pipeline yang menerapkan transformasi yang sesuai (misalnya power transform untuk kolom menceng dan scaling untuk penyeragaman rentang), lalu bandingkan satu model berbasis jarak dengan satu model berbasis pohon.

Luaran: kode pipeline, akurasi CV kedua model, dan 2-3 kalimat analisis kapan transformasi membantu.

Kriteria penilaian: (a) transformasi berada di dalam pipeline; (b) minimal dua keluarga model dibandingkan; (c) analisis menghubungkan hasil dengan sifat model.

# DATA AWAL (jangan diubah) - fitur menceng dan multi-skala.
Xr, yr = make_classification(n_samples=1200, n_features=6, n_informative=4,
                             n_redundant=0, random_state=7)
Xr[:, 1] = np.exp(Xr[:, 1])        # kolom menceng
Xr[:, 3] = Xr[:, 3] * 800          # kolom skala besar
Xr[:, 5] = np.abs(Xr[:, 5]) ** 3   # ekor panjang
print('Data:', Xr.shape, '| target seimbang:', np.bincount(yr))

Data: (1200, 6) | target seimbang: [598 602]

# Kerjakan di sini.
# Petunjuk: sklearn.preprocessing.PowerTransformer + StandardScaler di dalam Pipeline.