Proyek Akhir (Capstone)

🎬 Animasi: 3 Model, 3 Boundary

Data sama, 3 model → 3 boundary berbeda (LogReg lurus, KNN halus, Tree kotak-kotak). Tidak ada model ‘terbaik’ universal.

📌 Proyek 1: Prediksi Churn Pelanggan

Bisnis: pelanggan dengan tenure pendek dan tagihan tinggi cenderung pergi. Bangun model klasifikasi dan evaluasi dengan metrik yang tepat.

print("=== PROYEK 1: PREDIKSI CHURN ===\n")
random_seed(42)
n = 150
X, y = [], []
for _ in range(n):
    tenure = randint(1, 70)              # bulan
    charge = 20.0 + 60.0 * rand()        # tagihan/bulan
    age = 18 + randint(0, 50)
    # churn makin mungkin jika tenure pendek & charge tinggi
    z = -0.06 * tenure + 0.03 * (charge - 50.0)
    p_churn = sigmoid(z)
    X.append([tenure, charge, age])
    y.append(1 if rand() < p_churn else 0)

Xtr, Xte, ytr, yte = train_test_split(X, y, test_size=0.3)

# Standardisasi (dari TRAIN saja!)
mu = [mean([r[j] for r in Xtr]) for j in range(3)]
sd = [max(std([r[j] for r in Xtr]), 1e-6) for j in range(3)]
def scale(rows):
    return [[(r[j] - mu[j]) / sd[j] for j in range(3)] for r in rows]
Xtr_s, Xte_s = scale(Xtr), scale(Xte)

# Logistic regression
w = [0.0] * 4
Xb = add_bias(Xtr_s)
for epoch in range(300):
    for i in range(len(Xb)):
        p = sigmoid(dot(w, Xb[i]))
        w = vec_sub(w, vec_scale(Xb[i], 0.5 * (p - ytr[i])))

pred = [1 if sigmoid(dot(w, [1.0] + r)) >= 0.5 else 0 for r in Xte_s]
proba = [sigmoid(dot(w, [1.0] + r)) for r in Xte_s]

print("Churn di test: %d dari %d" % (yte.count(1), len(yte)))
print("Accuracy  = %.1f%%" % (accuracy(yte, pred) * 100))
print("Precision = %.2f  (yang diprediksi churn, berapa benar churn?)"
      % precision(yte, pred))
print("Recall    = %.2f  (pelanggan churn, berapa tertangkap?)"
      % recall(yte, pred))
print("F1        = %.2f" % f1_score(yte, pred))
print("\n→ Untuk retensi pelanggan, RECALL penting:")
print("  jangan sampai pelanggan yang akan pergi tidak terdeteksi.")

Output:

📌 Proyek 2: Segmentasi Pelanggan (K-Means)

Bisnis: bagi pelanggan menjadi segmen berdasarkan pendapatan, pengeluaran, dan usia — untuk marketing yang tepat sasaran.

print("=== PROYEK 2: SEGMENTASI PELANGGAN (K-MEANS 3D) ===\n")
random_seed(42)
# 3 segmen "alami": (pendapatan, pengeluaran, usia)
centers = [(30, 20, 25), (80, 60, 45), (150, 120, 35)]
X, true = [], []
for ci, (c1, c2, c3) in enumerate(centers):
    for _ in range(40):
        X.append([c1 + 8 * randn(), c2 + 8 * randn(), c3 + 5 * randn()])
        true.append(ci)

# K-means (3D)
k = 3
random_seed(1)
idx = shuffle(list(range(len(X))))[:k]
cents = [X[i][:] for i in idx]
for it in range(50):
    labels = [argmin([euclidean(p, c) for c in cents]) for p in X]
    new_c = []
    for ci in range(k):
        mem = [p for p, l in zip(X, labels) if l == ci]
        new_c.append([mean([q[j] for q in mem]) for j in range(3)] if mem else cents[ci])
    moved = sum(euclidean(a, b) for a, b in zip(cents, new_c))
    cents = new_c
    if moved < 0.01:
        break
labels = [argmin([euclidean(p, c) for c in cents]) for p in X]

print("Profil tiap segmen (mean pendapatan, pengeluaran, usia):")
for ci in range(k):
    mem = [p for p, l in zip(X, labels) if l == ci]
    prof = [mean([q[j] for q in mem]) for j in range(3)]
    print("  Segmen %d (n=%2d): pendapatan=%.0f  pengeluaran=%.0f  usia=%.0f"
          % (ci, len(mem), prof[0], prof[1], prof[2]))

# Cocokkan segmen dengan "sebenarnya" (untuk verifikasi)
print("\nKomposisi segmen (label asli 0/1/2):")
for ci in range(k):
    idxs = [i for i, l in enumerate(labels) if l == ci]
    comp = [sum(1 for i in idxs if true[i] == c) for c in range(3)]
    print("  Segmen %d: asli0=%d asli1=%d asli2=%d" % (ci, comp[0], comp[1], comp[2]))
print("→ Tiap segmen seharusnya didominasi SATU label asli.")

Output:

📌 Proyek 3: Deteksi Fraud (Data Tidak Seimbang)

Bisnis: transaksi fraud langka (~5%). Accuracy bukan metrik yang tepat — fokus ke recall (jangan ada fraud yang lolos).

print("=== PROYEK 3: DETEKSI FRAUD (5% POSITIF) ===\n")
random_seed(42)
X, y = [], []
for _ in range(950):   # normal
    X.append([2 + 10 * rand(), 1 + 5 * rand()])
    y.append(0)
for _ in range(50):    # fraud: nominal & frekuensi tidak wajar
    X.append([20 + 30 * rand(), 8 + 10 * rand()])
    y.append(1)

Xtr, Xte, ytr, yte = train_test_split(X, y, test_size=0.3)

# KNN (k=5)
def knn_predict(x, k=5):
    dists = sorted((euclidean(x, xt), yt) for xt, yt in zip(Xtr, ytr))[:k]
    votes = {}
    for d, l in dists:
        votes[l] = votes.get(l, 0) + 1
    return max(votes, key=votes.get)

pred = [knn_predict(x) for x in Xte]

print("Fraud di test: %d dari %d" % (yte.count(1), len(yte)))
print("Accuracy = %.1f%%  ← TIDAK boleh jadi tolok ukur!"
      % (accuracy(yte, pred) * 100))
print("Recall   = %.2f  ← fraud yang tertangkap (yang PALING penting)"
      % recall(yte, pred))
print("Precision= %.2f  ← alarm yang benar-benar fraud"
      % precision(yte, pred))

# Model bodoh: prediksi semua normal
pred0 = [0] * len(yte)
print("\nModel 'prediksi semua normal':")
print("  Accuracy = %.1f%%  tapi Recall = %.2f (semua fraud lolos!)"
      % (accuracy(yte, pred0) * 100, recall(yte, pred0)))
print("→ Di fraud detection, kita rela banyak FALSE POSITIVE")
print("  (transaksi normal ditahan untuk review) demi RECALL tinggi.")

Output:

📌 Proyek 4: Perbandingan Model End-to-End

Alur lengkap: data → split → 3 model berbeda → evaluasi → kesimpulan. Ini pola kerja data scientist sehari-hari.

print("=== PROYEK 4: BANDINGKAN 3 MODEL (MOONS) ===\n")
random_seed(42)
X, y = make_moons(200, noise=0.2)
Xtr, Xte, ytr, yte = train_test_split(X, y, test_size=0.3)

# ── Model 1: Logistic Regression ──
w = [0.0, 0.0, 0.0]
Xb = add_bias(Xtr)
for epoch in range(300):
    for i in range(len(Xb)):
        p = sigmoid(dot(w, Xb[i]))
        w = vec_sub(w, vec_scale(Xb[i], 0.5 * (p - ytr[i])))
m1 = [1 if sigmoid(dot(w, [1.0] + x)) >= 0.5 else 0 for x in Xte]

# ── Model 2: KNN (k=7) ──
def knn(x, k=7):
    d = sorted((euclidean(x, xt), yt) for xt, yt in zip(Xtr, ytr))[:k]
    v = {}
    for dd, l in d:
        v[l] = v.get(l, 0) + 1
    return max(v, key=v.get)
m2 = [knn(x) for x in Xte]

# ── Model 3: Decision Tree (depth 3) ──
def entropy(yy):
    n = len(yy); h = 0.0
    for l in set(yy):
        p = sum(1 for v in yy if v == l) / n
        h -= p * math.log2(p)
    return h
def best_split(XX, yy):
    ph = entropy(yy); n = len(yy); best = None
    for f in range(len(XX[0])):
        vals = sorted(set(r[f] for r in XX))
        for a, b in zip(vals, vals[1:]):
            thr = (a + b) / 2.0
            yl = [yy[i] for i in range(n) if XX[i][f] <= thr]
            yr = [yy[i] for i in range(n) if XX[i][f] > thr]
            if not yl or not yr:
                continue
            hc = (len(yl) * entropy(yl) + len(yr) * entropy(yr)) / n
            g = ph - hc
            if best is None or g > best[0]:
                best = (g, f, thr)
    return best
def build_tree(XX, yy, d, md):
    votes = {}
    for l in yy:
        votes[l] = votes.get(l, 0) + 1
    maj = max(votes, key=votes.get)
    if d >= md or len(set(yy)) == 1:
        return ("leaf", maj)
    s = best_split(XX, yy)
    if s is None or s[0] <= 1e-9:
        return ("leaf", maj)
    g, f, thr = s
    li = [i for i in range(len(XX)) if XX[i][f] <= thr]
    ri = [i for i in range(len(XX)) if XX[i][f] > thr]
    return ("split", f, thr,
            build_tree([XX[i] for i in li], [yy[i] for i in li], d + 1, md),
            build_tree([XX[i] for i in ri], [yy[i] for i in ri], d + 1, md))
def predict(t, x):
    if t[0] == "leaf":
        return t[1]
    _, f, thr, l, r = t
    return predict(l if x[f] <= thr else r, x)
tree = build_tree(Xtr, ytr, 0, 3)
m3 = [predict(tree, x) for x in Xte]

# ── Laporan ──
print_table(["Model", "Accuracy", "Precision", "Recall", "F1"],
            ["logistic reg.", "%.1f%%" % (accuracy(yte, m1) * 100),
             "%.2f" % precision(yte, m1), "%.2f" % recall(yte, m1),
             "%.2f" % f1_score(yte, m1)],
            ["KNN (k=7)",     "%.1f%%" % (accuracy(yte, m2) * 100),
             "%.2f" % precision(yte, m2), "%.2f" % recall(yte, m2),
             "%.2f" % f1_score(yte, m2)],
            ["decision tree", "%.1f%%" % (accuracy(yte, m3) * 100),
             "%.2f" % precision(yte, m3), "%.2f" % recall(yte, m3),
             "%.2f" % f1_score(yte, m3)])
print("""
Kesimpulan:
• Data moons NON-LINEAR → logistic regression (garis lurus) tertinggal.
• KNN & tree bisa mengikuti bentuk melengkung.
• Di dunia nyata: coba 5-10 model, ukur di test set yang sama,
  lalu pilih berdasarkan metrik yang sesuai BISI (bukan accuracy buta).""")

🏁 Proyek Akhir (Capstone)

🎬 Animasi: 3 Model, 3 Boundary

📌 Proyek 1: Prediksi Churn Pelanggan

Output:

📌 Proyek 2: Segmentasi Pelanggan (K-Means)

Output:

📌 Proyek 3: Deteksi Fraud (Data Tidak Seimbang)

Output:

📌 Proyek 4: Perbandingan Model End-to-End

Output: