from pathlib import Path
import json
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
plt.style.use('seaborn-v0_8-whitegrid')
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 cacheNotebook Bab 15 - Embedding dan Model Tabular
Notebook Bab 15 ini punya dua bagian. Bagian Demo tinggal Anda jalankan lalu amati keluarannya; bagian Mini Project berisi soal dan data yang Anda kerjakan sendiri.
Demo tidak memperkenalkan data teks/citra/audio baru. Ia membaca ringkasan terverifikasi dari Bab 11-12, lalu memakai subset UCI Covertype 50k untuk membandingkan model tabular pada skala yang lebih besar.
Persiapan
Section 1 - Demo: Reuse Embedding dan Skala Tabular
Reuse dari Bab 11-12
REUSE_DIR = section_data_dir('ch15_sms_fashion_fsdd_reuse')
COV_DIR = section_data_dir('ch15_covtype_50k')
reuse_stats = json.loads((REUSE_DIR / 'verified_stats.json').read_text(encoding='utf-8'))
reuse = pd.read_csv(REUSE_DIR / 'reuse_probe_scores.csv')
print(f"Reuse rows: {reuse_stats['rows']} from {reuse_stats['referenced_snapshots']}")
print(reuse.to_string(index=False, float_format=lambda x: f'{x:.4f}'))
plot_reuse = reuse.copy()
plot_reuse['label'] = plot_reuse['source'] + ' | ' + plot_reuse['representation'] + ' | ' + plot_reuse['metric']
fig, ax = plt.subplots(figsize=(8, 3.8))
ax.barh(plot_reuse['label'], plot_reuse['score'], color='#4c78a8')
ax.set_xlim(0, 1)
ax.set_xlabel('Score')
ax.set_title('Ringkasan probe dari bab sebelumnya')
for i, value in enumerate(plot_reuse['score']):
ax.text(value + 0.01, i, f'{value:.3f}', va='center')
plt.tight_layout()
plt.show()Reuse rows: 6 from ['data/section1/ch11_sms_spam', 'data/section1/ch12_fashion_mnist', 'data/section1/ch12_fsdd']
source modality representation metric score
SMS Spam text char_wb TF-IDF average_precision 0.9858
SMS Spam text char_wb TF-IDF f1 0.9648
Fashion-MNIST image HOG accuracy 0.8090
Fashion-MNIST image PCA pixel embedding accuracy 0.8300
FSDD audio MFCC accuracy 0.9500
FSDD audio log-mel PCA accuracy 0.8933
Covertype: tabular besar dengan fitur campuran
cov_stats = json.loads((COV_DIR / 'verified_stats.json').read_text(encoding='utf-8'))
covtype = pd.read_parquet(COV_DIR / 'covtype_50k.parquet')
class_counts = pd.read_csv(COV_DIR / 'covtype_50k_class_counts.csv')
model_scores = pd.read_csv(COV_DIR / 'covtype_model_scores.csv')
print(f"Full rows: {cov_stats['full_rows']:,}; subset rows: {cov_stats['sample_rows']:,}")
print(f"Features: {cov_stats['features']} ({cov_stats['quantitative_columns']} quantitative + {cov_stats['binary_wilderness_soil_columns']} binary wilderness/soil)")
print(f"Missing values: {cov_stats['missing_values']}")
print(model_scores.to_string(index=False, float_format=lambda x: f'{x:.4f}'))
fig, axes = plt.subplots(1, 2, figsize=(10, 3.8))
axes[0].bar(class_counts['cover_type'].astype(str), class_counts['sample_count'], color='#72b7b2')
axes[0].set_title('Class imbalance pada subset 50k')
axes[0].set_xlabel('Cover_Type')
axes[0].set_ylabel('Rows')
model_plot = model_scores.melt(id_vars='model', value_vars=['accuracy','macro_f1'], var_name='metric', value_name='score')
for metric, group in model_plot.groupby('metric'):
offset = -0.18 if metric == 'accuracy' else 0.18
x = np.arange(len(group)) + offset
axes[1].bar(x, group['score'], width=0.35, label=metric)
axes[1].set_xticks(np.arange(len(model_scores)), model_scores['model'], rotation=10)
axes[1].set_ylim(0, 1)
axes[1].set_title('GBDT vs MLP kecil pada tabular')
axes[1].legend()
plt.tight_layout()
plt.show()Full rows: 581,012; subset rows: 50,000
Features: 54 (10 quantitative + 44 binary wilderness/soil)
Missing values: 0
model accuracy macro_f1 train_rows test_rows
HistGradientBoosting 0.8440 0.7783 35000 15000
MLP_scaled 0.7531 0.5307 35000 15000
?? Amati. Skor reuse dari Bab 11-12 menunjukkan bahwa embedding atau fitur ringkas bisa sangat berguna pada teks, citra, dan audio. Tetapi Covertype mengingatkan bahwa tabular besar punya aturan main sendiri: pada subset 50k ini, gradient boosting sederhana mengalahkan MLP kecil, terutama pada macro-F1. Ini bukan vonis anti-deep-learning; ini bukti bahwa model tabular tetap perlu diuji dengan ruler lokal.
Section 2 - Mini Project
Soal
Gunakan subset Covertype untuk menguji keputusan representasi tabular.
Tugas:
- Pakai
mini_covtypedanfeature_colsdari sel di bawah. - Bandingkan model yang memakai semua fitur dengan model yang hanya memakai 10 fitur kuantitatif.
- Laporkan accuracy dan macro-F1, lalu jelaskan apa kontribusi fitur wilderness/soil biner.
Luaran: tabel metrik dan 3-4 kalimat analisis.
Kriteria penilaian: (a) split stratified; (b) Cover_Type tidak masuk fitur; (c) macro-F1 ikut dilaporkan karena kelas tidak seimbang.
# DATA AWAL (jangan diubah)
mini_covtype = covtype.sample(n=12000, random_state=42).reset_index(drop=True)
feature_cols = [c for c in mini_covtype.columns if c != 'Cover_Type']
quantitative_cols = feature_cols[:10]
binary_cols = feature_cols[10:]
print('Mini shape:', mini_covtype.shape)
print('Quantitative:', len(quantitative_cols), '| binary:', len(binary_cols))
mini_covtype[['Elevation','Aspect','Slope','Cover_Type']].head()Mini shape: (12000, 55)
Quantitative: 10 | binary: 44| Elevation | Aspect | Slope | Cover_Type | |
|---|---|---|---|---|
| 0 | 3076 | 115 | 15 | 2 |
| 1 | 2786 | 70 | 25 | 2 |
| 2 | 2838 | 231 | 12 | 5 |
| 3 | 3014 | 350 | 5 | 2 |
| 4 | 3254 | 251 | 16 | 1 |
# Kerjakan di sini.
# Petunjuk: pakai train_test_split(..., stratify=mini_covtype['Cover_Type']); bandingkan semua fitur vs quantitative_cols.