15.11-dars: Belgi muhimligi

MDI (Mean Decrease in Impurity) = feature_importances_

  har belgi uchun: u bergan nopoklik kamayishlarining
  og'irlikli yig'indisi (barcha daraxtlar bo'yicha)

UCH TUZOQ:
  1. KARDINALLIK moyilligi — ko'p noyob qiymatli belgi ustun
     (ID, vaqt belgisi, uzluksiz belgi > kategoriya)
  2. O'QUV ma'lumotida hisoblanadi — overfitting aks etadi
  3. KORRELYATSIYALI belgilar orasida tasodifiy taqsimlanadi

sklearn hujjati ham ogohlantiradi: permutation importance ni tavsiya qiladi

G'oya: belgini ARALASHTIRSAK, sifat qanchaga tushadi?

  1. Asosiy ball hisoblanadi (validatsiya/test to'plamida)
  2. Bitta belgi ustuni tasodifiy aralashtiriladi
  3. Ball qayta hisoblanadi
  4. muhimlik = asosiy - aralashtirilgan
  5. n_repeats marta takrorlanadi -> o'rtacha va std

sklearn:
  from sklearn.inspection import permutation_importance
  r = permutation_importance(model, Xte, yte, n_repeats=20,
                             scoring="roc_auc", random_state=0)
  r.importances_mean, r.importances_std

Eng halol, lekin eng qimmat usul:

  har belgi uchun: uni OLIB TASHLAB, modelni QAYTA o'qitish
  muhimlik = to'liq model balli - belgisiz model balli

  + haqiqiy hissani o'lchaydi
  - p ta qayta o'qitish kerak
  - korrelyatsiyali belgilarda ikkalasi ham "muhimsiz" chiqadi

Amalda: faqat kam belgili yoki muhim qarorlarda

SHAP (SHapley Additive exPlanations) — o'yin nazariyasidan

  har BASHORAT uchun har belgining hissasi:
    f(x) = asosiy_qiymat + sum(shap_qiymat_i)

  + lokal (bitta bashorat) va global (o'rtacha |SHAP|) tushuntirish
  + daraxtlar uchun aniq va tez algoritm (TreeSHAP)
  + yo'nalish ko'rinadi (musbat/manfiy ta'sir)
  - alohida kutubxona (shap)
  - korrelyatsiyali belgilarda talqin murakkab

sklearn muqobili: partial_dependence, PartialDependenceDisplay

Muammo: a va b korrelyatsiyasi 0.95
  MDI: muhimlik ikkiga bo'linadi (ikkalasi ham "o'rtacha")
  Permutation: a ni aralashtirsak, model b dan foydalanadi
               -> ikkalasi ham "muhimsiz" chiqadi

YECHIMLAR:
  1. Ierarxik klasterlash bilan guruhlash, guruhni birga aralashtirish
  2. Korrelyatsiyali guruhdan bittasini qoldirish
  3. Guruh darajasida muhimlik hisoblash
  4. Conditional permutation importance

sklearn misoli: scipy.cluster.hierarchy bilan Spearman asosida

Muhimlik BASHORAT uchun foydalilikni o'lchaydi, SABABNI emas

Misol: "kasalxonada yotgan kunlar soni" o'lim ehtimolini yaxshi bashorat qiladi
  -> lekin kunlarni kamaytirish o'limni kamaytirmaydi

Sababiy xulosa uchun kerak:
  - eksperiment (A/B test)
  - sababiy modellar (DAG, instrumental o'zgaruvchilar)
  - domen bilimi

"Bu belgi muhim -> uni o'zgartiraylik" - ENG KENG TARQALGAN XATO

from sklearn.inspection import permutation_importance, partial_dependence

model.feature_importances_             # MDI - ehtiyot bo'ling

r = permutation_importance(model, Xte, yte, n_repeats=20,
                           scoring="roc_auc", random_state=0, n_jobs=-1)
tartib = np.argsort(-r.importances_mean)
for i in tartib[:10]:
    print(nomlar[i], r.importances_mean[i], r.importances_std[i])

# korrelyatsiyali guruhlar
from scipy.cluster.hierarchy import fcluster, linkage
from scipy.stats import spearmanr
korr = spearmanr(X).correlation
guruh = fcluster(linkage(1 - np.abs(korr), "average"), 0.3, "distance")
QOIDA: permutation ni TEST da · guruhlab aralashtir ·
       sababiy talqin qilma

MDI: tez, kardinallikka moyil, o'quvda, korrelyatsiyada chalg'itadi
Permutation: test da hisobla, n_repeats >= 10
Drop-column: eng halol, p ta qayta o'qitish
SHAP: lokal + global, yo'nalish ko'rinadi
Muhimlik != sababiylik

"""Tasodifiy ID nega "eng muhim" chiqadi (real numpy/pandas/sklearn)."""

import numpy as np
import pandas as pd
from sklearn.ensemble import RandomForestClassifier
from sklearn.inspection import permutation_importance
from sklearn.model_selection import train_test_split


def yarat(seed: int = 5, n: int = 4000) -> pd.DataFrame:
    rng = np.random.default_rng(seed)
    haqiqiy1 = rng.normal(0, 1, n)
    haqiqiy2 = rng.normal(0, 1, n)
    ikkilik = rng.integers(0, 2, n).astype(float)        # 2 daraja
    uch_daraja = rng.integers(0, 3, n).astype(float)     # 3 daraja
    tasodifiy_id = rng.permutation(n).astype(float)      # n daraja, ma'nosiz
    tasodifiy_son = rng.normal(0, 1, n)                  # uzluksiz, ma'nosiz
    kuch = 1.5 * haqiqiy1 - 1.2 * haqiqiy2 + 0.8 * ikkilik
    y = (rng.random(n) < 1 / (1 + np.exp(-kuch))).astype(int)
    return pd.DataFrame({"haqiqiy1": haqiqiy1, "haqiqiy2": haqiqiy2,
                         "ikkilik": ikkilik, "uch_daraja": uch_daraja,
                         "tasodifiy_id": tasodifiy_id,
                         "tasodifiy_son": tasodifiy_son, "y": y})


def main() -> None:
    df = yarat()
    nomlar = [c for c in df.columns if c != "y"]
    X, y = df[nomlar], df["y"]
    Xtr, Xte, ytr, yte = train_test_split(X, y, test_size=0.3, random_state=0,
                                          stratify=y)
    o = RandomForestClassifier(n_estimators=300, random_state=0,
                               n_jobs=1).fit(Xtr, ytr)

    print("=== 1. Belgilarning noyob qiymatlari ===")
    for nom in nomlar:
        print(f"  {nom:<15}: {X[nom].nunique():>5} noyob qiymat")

    print("\n=== 2. MDI (feature_importances_) ===")
    mdi = o.feature_importances_
    for i in np.argsort(-mdi):
        print(f"  {nomlar[i]:<15}: {mdi[i]:.4f}")
    print(f"  (haqiqiy signal faqat: haqiqiy1, haqiqiy2, ikkilik)")

    print("\n=== 3. Permutation importance (test to'plamida) ===")
    r = permutation_importance(o, Xte, yte, n_repeats=20, scoring="roc_auc",
                               random_state=0, n_jobs=1)
    for i in np.argsort(-r.importances_mean):
        print(f"  {nomlar[i]:<15}: {r.importances_mean[i]:>+8.4f} "
              f"(std {r.importances_std[i]:.4f})")

    print("\n=== 4. Permutation: o'quv va test farqi ===")
    r_tr = permutation_importance(o, Xtr, ytr, n_repeats=10,
                                  scoring="roc_auc", random_state=0, n_jobs=1)
    print(f"  {'belgi':<15} {'MDI':>8} {'perm(o_quv)':>13} "
          f"{'perm(test)':>12}")
    for i in range(len(nomlar)):
        print(f"  {nomlar[i]:<15} {mdi[i]:>8.4f} "
              f"{r_tr.importances_mean[i]:>+13.4f} "
              f"{r.importances_mean[i]:>+12.4f}")
    print("  ⭐ MDI ma'nosiz belgini yuqori baholaydi")


if __name__ == "__main__":
    main()

=== 1. Belgilarning noyob qiymatlari ===
  haqiqiy1       :  4000 noyob qiymat
  haqiqiy2       :  4000 noyob qiymat
  ikkilik        :     2 noyob qiymat
  uch_daraja     :     3 noyob qiymat
  tasodifiy_id   :  4000 noyob qiymat
  tasodifiy_son  :  4000 noyob qiymat

=== 2. MDI (feature_importances_) ===
  haqiqiy1       : 0.3624
  haqiqiy2       : 0.2930
  tasodifiy_id   : 0.1498
  tasodifiy_son  : 0.1460
  uch_daraja     : 0.0302
  ikkilik        : 0.0186
  (haqiqiy signal faqat: haqiqiy1, haqiqiy2, ikkilik)

=== 3. Permutation importance (test to'plamida) ===
  haqiqiy1       :  +0.2055 (std 0.0146)
  haqiqiy2       :  +0.1301 (std 0.0105)
  ikkilik        :  +0.0092 (std 0.0028)
  tasodifiy_son  :  +0.0030 (std 0.0022)
  uch_daraja     :  +0.0005 (std 0.0018)
  tasodifiy_id   :  -0.0021 (std 0.0028)

=== 4. Permutation: o'quv va test farqi ===
  belgi                MDI   perm(o_quv)   perm(test)
  haqiqiy1          0.3624       +0.2792      +0.2055
  haqiqiy2          0.2930       +0.2110      +0.1301
  ikkilik           0.0186       +0.0291      +0.0092
  uch_daraja        0.0302       +0.0048      +0.0005
  tasodifiy_id      0.1498       +0.0285      -0.0021
  tasodifiy_son     0.1460       +0.0231      +0.0030
  ⭐ MDI ma'nosiz belgini yuqori baholaydi

"""Muhimlik nega "yo'qoladi" va uni qanday tiklash (real numpy/scipy/sklearn)."""

import numpy as np
from scipy.cluster.hierarchy import fcluster, linkage
from scipy.spatial.distance import squareform
from scipy.stats import spearmanr
from sklearn.ensemble import RandomForestClassifier
from sklearn.inspection import permutation_importance
from sklearn.metrics import roc_auc_score
from sklearn.model_selection import train_test_split


def yarat(seed: int = 8, n: int = 4000):
    rng = np.random.default_rng(seed)
    asos = rng.normal(0, 1, n)
    # uchta korrelyatsiyali nusxa
    a1 = asos + rng.normal(0, 0.25, n)
    a2 = asos + rng.normal(0, 0.25, n)
    a3 = asos + rng.normal(0, 0.25, n)
    mustaqil = rng.normal(0, 1, n)
    shovqin = rng.normal(0, 1, (n, 3))
    X = np.column_stack([a1, a2, a3, mustaqil, shovqin])
    kuch = 1.6 * asos + 1.2 * mustaqil
    y = (rng.random(n) < 1 / (1 + np.exp(-kuch))).astype(int)
    nomlar = ["nusxa1", "nusxa2", "nusxa3", "mustaqil", "shovqin1",
              "shovqin2", "shovqin3"]
    return X, y, nomlar


def main() -> None:
    X, y, nomlar = yarat()
    Xtr, Xte, ytr, yte = train_test_split(X, y, test_size=0.3, random_state=0,
                                          stratify=y)
    o = RandomForestClassifier(n_estimators=300, random_state=0,
                               n_jobs=1).fit(Xtr, ytr)

    print("=== 1. Korrelyatsiya ===")
    print(f"  nusxa1 - nusxa2: {np.corrcoef(X[:, 0], X[:, 1])[0, 1]:.4f}")
    print(f"  nusxa1 - mustaqil: {np.corrcoef(X[:, 0], X[:, 3])[0, 1]:.4f}")

    print("\n=== 2. MDI va permutation ===")
    mdi = o.feature_importances_
    r = permutation_importance(o, Xte, yte, n_repeats=20, scoring="roc_auc",
                               random_state=0, n_jobs=1)
    print(f"  {'belgi':<12} {'MDI':>8} {'permutation':>13}")
    for i in range(len(nomlar)):
        print(f"  {nomlar[i]:<12} {mdi[i]:>8.4f} "
              f"{r.importances_mean[i]:>+13.4f}")
    print("  (uchta nusxa birgalikda mustaqildan kuchli, lekin alohida past)")

    print("\n=== 3. Ierarxik klasterlash ===")
    korr = spearmanr(X).correlation
    masofa = 1 - np.abs(korr)
    np.fill_diagonal(masofa, 0.0)
    masofa = (masofa + masofa.T) / 2
    Z = linkage(squareform(masofa, checks=False), "average")
    guruhlar = fcluster(Z, 0.3, criterion="distance")
    for g in np.unique(guruhlar):
        azolar = [nomlar[i] for i in range(len(nomlar)) if guruhlar[i] == g]
        print(f"  guruh {g}: {azolar}")

    print("\n=== 4. Guruh bo'yicha aralashtirish ===")
    asosiy = roc_auc_score(yte, o.predict_proba(Xte)[:, 1])
    rng = np.random.default_rng(0)
    print(f"  asosiy AUC: {asosiy:.4f}")
    print(f"  {'guruh':<28} {'muhimlik':>10}")
    for g in np.unique(guruhlar):
        indekslar = [i for i in range(len(nomlar)) if guruhlar[i] == g]
        ballar = []
        for _ in range(15):
            Xa = Xte.copy()
            aralash = rng.permutation(len(Xa))
            Xa[:, indekslar] = Xa[aralash][:, indekslar]   # BIRGA aralashtirish
            ballar.append(roc_auc_score(yte, o.predict_proba(Xa)[:, 1]))
        nom = ",".join(nomlar[i] for i in indekslar)
        print(f"  {nom:<28} {asosiy - np.mean(ballar):>+10.4f}")
    print("  ⭐ Guruhlab aralashtirish haqiqiy hissani ko'rsatadi")


if __name__ == "__main__":
    main()

=== 1. Korrelyatsiya ===
  nusxa1 - nusxa2: 0.9445
  nusxa1 - mustaqil: -0.0070

=== 2. MDI va permutation ===
  belgi             MDI   permutation
  nusxa1         0.1753       +0.0476
  nusxa2         0.1575       +0.0224
  nusxa3         0.1739       +0.0233
  mustaqil       0.2170       +0.1210
  shovqin1       0.0938       +0.0000
  shovqin2       0.0924       -0.0010
  shovqin3       0.0901       -0.0008
  (uchta nusxa birgalikda mustaqildan kuchli, lekin alohida past)

=== 3. Ierarxik klasterlash ===
  guruh 1: ['nusxa1', 'nusxa2', 'nusxa3']
  guruh 2: ['shovqin1']
  guruh 3: ['mustaqil']
  guruh 4: ['shovqin3']
  guruh 5: ['shovqin2']

=== 4. Guruh bo'yicha aralashtirish ===
  asosiy AUC: 0.8442
  guruh                          muhimlik
  nusxa1,nusxa2,nusxa3            +0.2476
  shovqin1                        -0.0005
  mustaqil                        +0.1217
  shovqin3                        -0.0004
  shovqin2                        -0.0027
  ⭐ Guruhlab aralashtirish haqiqiy hissani ko'rsatadi

"""Eng halol usul va ta'sir yo'nalishi (real numpy/sklearn)."""

import numpy as np
from sklearn.ensemble import HistGradientBoostingClassifier
from sklearn.inspection import partial_dependence, permutation_importance
from sklearn.metrics import roc_auc_score
from sklearn.model_selection import train_test_split


def yarat(seed: int = 13, n: int = 5000):
    rng = np.random.default_rng(seed)
    X = rng.normal(0, 1, (n, 6))
    kuch = (1.5 * X[:, 0] - 1.2 * X[:, 1]
            + 1.0 * ((X[:, 2] > 0.5) & (X[:, 3] > 0))
            + 0.4 * X[:, 4])
    y = (rng.random(n) < 1 / (1 + np.exp(-kuch))).astype(int)
    nomlar = ["kuchli+", "kuchli-", "ozaro_a", "ozaro_b", "zaif", "shovqin"]
    return X, y, nomlar


def main() -> None:
    X, y, nomlar = yarat()
    Xtr, Xte, ytr, yte = train_test_split(X, y, test_size=0.3, random_state=0,
                                          stratify=y)

    def oqit(Xa, Xb):
        m = HistGradientBoostingClassifier(learning_rate=0.1, max_iter=300,
                                           early_stopping=True,
                                           validation_fraction=0.15,
                                           n_iter_no_change=20,
                                           random_state=0).fit(Xa, ytr)
        return m, roc_auc_score(yte, m.predict_proba(Xb)[:, 1])

    print("=== 1. To'liq model ===")
    model, toliq = oqit(Xtr, Xte)
    print(f"  test AUC: {toliq:.4f}")

    print("\n=== 2. Drop-column muhimligi ===")
    print(f"  {'olib tashlangan':<18} {'AUC':>9} {'yo_qotish':>11}")
    drop = {}
    for i, nom in enumerate(nomlar):
        qolgan = [j for j in range(len(nomlar)) if j != i]
        _, a = oqit(Xtr[:, qolgan], Xte[:, qolgan])
        drop[nom] = toliq - a
        print(f"  {nom:<18} {a:>9.4f} {toliq - a:>+11.4f}")

    print("\n=== 3. Uch usul yonma-yon ===")
    r = permutation_importance(model, Xte, yte, n_repeats=20,
                               scoring="roc_auc", random_state=0, n_jobs=1)
    print(f"  {'belgi':<12} {'permutation':>13} {'drop-column':>13}")
    for i, nom in enumerate(nomlar):
        print(f"  {nom:<12} {r.importances_mean[i]:>+13.4f} "
              f"{drop[nom]:>+13.4f}")

    print("\n=== 4. Ta'sir yo'nalishi (partial dependence) ===")
    for i in [0, 1, 4]:
        pd_natija = partial_dependence(model, Xte, [i], grid_resolution=5,
                                       kind="average")
        qiymatlar = pd_natija["grid_values"][0]
        ortacha = pd_natija["average"][0]
        print(f"  {nomlar[i]:<10}: x = {np.round(qiymatlar, 2).tolist()}")
        print(f"  {'':<10}  f = {np.round(ortacha, 3).tolist()}")
    print("  ⭐ Muhimlik kattaligini, PD yo'nalishini ko'rsatadi")


if __name__ == "__main__":
    main()

=== 1. To'liq model ===
  test AUC: 0.8300

=== 2. Drop-column muhimligi ===
  olib tashlangan          AUC   yo_qotish
  kuchli+               0.6880     +0.1420
  kuchli-               0.7632     +0.0668
  ozaro_a               0.8287     +0.0013
  ozaro_b               0.8208     +0.0092
  zaif                  0.8211     +0.0089
  shovqin               0.8349     -0.0049

=== 3. Uch usul yonma-yon ===
  belgi          permutation   drop-column
  kuchli+            +0.1986       +0.1420
  kuchli-            +0.1228       +0.0668
  ozaro_a            +0.0073       +0.0013
  ozaro_b            +0.0108       +0.0092
  zaif               +0.0146       +0.0089
  shovqin            -0.0001       -0.0049

=== 4. Ta'sir yo'nalishi (partial dependence) ===
  kuchli+   : x = [-1.65, -0.82, 0.01, 0.84, 1.67]
              f = [-2.309, -1.149, 0.066, 1.621, 1.819]
  kuchli-   : x = [-1.75, -0.92, -0.09, 0.74, 1.57]
              f = [1.906, 1.247, 0.097, -0.694, -2.284]
  zaif      : x = [-1.62, -0.8, 0.01, 0.83, 1.65]
              f = [-0.358, -0.175, 0.094, 0.486, 0.586]
  ⭐ Muhimlik kattaligini, PD yo'nalishini ko'rsatadi

"""Bashorat kuchi va sababiy ta'sir farqi (real numpy/sklearn)."""

import numpy as np
from sklearn.ensemble import RandomForestRegressor
from sklearn.inspection import permutation_importance
from sklearn.linear_model import LinearRegression
from sklearn.metrics import r2_score
from sklearn.model_selection import train_test_split


def yarat(seed: int = 17, n: int = 5000):
    """Sababiy tuzilma: davolash -> tuzalish; kunlar <- og'irlik -> tuzalish."""
    rng = np.random.default_rng(seed)
    ogirlik = rng.normal(0, 1, n)                 # kasallik og'irligi (yashirin)
    davolash = (rng.random(n) < 0.5).astype(float)
    # kunlar og'irlikning NATIJASI, sababi emas
    kunlar = 5 + 3 * ogirlik - 1.0 * davolash + rng.normal(0, 0.8, n)
    tuzalish = (10 - 2.5 * ogirlik + 3.0 * davolash + rng.normal(0, 1.0, n))
    X = np.column_stack([davolash, kunlar, rng.normal(0, 1, (n, 2))])
    nomlar = ["davolash", "kunlar", "shovqin1", "shovqin2"]
    return X, tuzalish, nomlar, ogirlik


def main() -> None:
    X, y, nomlar, ogirlik = yarat()
    Xtr, Xte, ytr, yte = train_test_split(X, y, test_size=0.3, random_state=0)
    o = RandomForestRegressor(n_estimators=300, max_features=0.5,
                              random_state=0, n_jobs=1).fit(Xtr, ytr)

    print("=== 1. Model sifati ===")
    print(f"  test R^2: {r2_score(yte, o.predict(Xte)):.4f}")

    print("\n=== 2. Muhimlik ===")
    r = permutation_importance(o, Xte, yte, n_repeats=20, scoring="r2",
                               random_state=0, n_jobs=1)
    print(f"  {'belgi':<12} {'MDI':>8} {'permutation':>13}")
    for i in np.argsort(-r.importances_mean):
        print(f"  {nomlar[i]:<12} {o.feature_importances_[i]:>8.4f} "
              f"{r.importances_mean[i]:>+13.4f}")
    print("  (kunlar eng muhim ko'rinadi)")

    print("\n=== 3. Sababiy haqiqat ===")
    print("  Haqiqiy tuzilma:")
    print("    og'irlik -> kunlar      (og'irlik kunlarni oshiradi)")
    print("    og'irlik -> tuzalish    (og'irlik tuzalishni kamaytiradi)")
    print("    davolash -> tuzalish    (+3.0)")
    print("    davolash -> kunlar      (-1.0)")
    print("  Ya'ni kunlar tuzalishga TA'SIR QILMAYDI - ikkalasi ham")
    print("  og'irlikning natijasi (umumiy sabab)")

    print("\n=== 4. Aralashuv (intervention) simulyatsiyasi ===")
    # kunlarni sun'iy kamaytirish tuzalishni o'zgartiradimi?
    Xa = Xte.copy()
    Xa[:, 1] -= 2.0                     # "kunlarni 2 ga kamaytiramiz"
    print(f"  model bashorati (kunlar -2): "
          f"{o.predict(Xa).mean():.4f} (asl {o.predict(Xte).mean():.4f})")
    print(f"  model o'zgarish kutmoqda: "
          f"{o.predict(Xa).mean() - o.predict(Xte).mean():+.4f}")
    print("  Lekin haqiqiy mexanizmda kunlar tuzalishga ta'sir qilmaydi:")
    print("  haqiqiy o'zgarish = 0.0000")
    # to'g'ri sababiy baho: og'irlikni nazoratga olish
    Xc = np.column_stack([X[:, 0], ogirlik])
    ch = LinearRegression().fit(Xc, y)
    print(f"  davolashning sababiy ta'siri (og'irlik nazoratda): "
          f"{ch.coef_[0]:+.4f} (haqiqiy +3.0000)")
    chs = LinearRegression().fit(X[:, [0]], y)
    print(f"  og'irliksiz baho: {chs.coef_[0]:+.4f}")
    print("  ⭐ Muhimlik bashorat uchun, sababiylik uchun emas")


if __name__ == "__main__":
    main()

=== 1. Model sifati ===
  test R^2: 0.8349

=== 2. Muhimlik ===
  belgi             MDI   permutation
  kunlar         0.7140       +1.2293
  davolash       0.1749       +0.2539
  shovqin2       0.0559       -0.0006
  shovqin1       0.0552       -0.0038
  (kunlar eng muhim ko'rinadi)

=== 3. Sababiy haqiqat ===
  Haqiqiy tuzilma:
    og'irlik -> kunlar      (og'irlik kunlarni oshiradi)
    og'irlik -> tuzalish    (og'irlik tuzalishni kamaytiradi)
    davolash -> tuzalish    (+3.0)
    davolash -> kunlar      (-1.0)
  Ya'ni kunlar tuzalishga TA'SIR QILMAYDI - ikkalasi ham
  og'irlikning natijasi (umumiy sabab)

=== 4. Aralashuv (intervention) simulyatsiyasi ===
  model bashorati (kunlar -2): 13.0663 (asl 11.5513)
  model o'zgarish kutmoqda: +1.5150
  Lekin haqiqiy mexanizmda kunlar tuzalishga ta'sir qilmaydi:
  haqiqiy o'zgarish = 0.0000
  davolashning sababiy ta'siri (og'irlik nazoratda): +3.0096 (haqiqiy +3.0000)
  og'irliksiz baho: +3.0667
  ⭐ Muhimlik bashorat uchun, sababiylik uchun emas

plt.barh(nomlar, model.feature_importances_)                      # ⚠️
r = permutation_importance(model, Xte, yte, n_repeats=20)         # ✅

permutation_importance(model, Xtr, ytr)                           # ⚠️
permutation_importance(model, Xte, yte)                           # ✅

# to'g'ridan-to'g'ri muhimlik ro'yxati                            # ⚠️
# avval spearmanr + klasterlash, keyin guruh bo'yicha             # ✅

# "kunlarni kamaytiramiz - natija yaxshilanadi"                   # ⚠️
# eksperiment yoki sababiy model kerak                            # ✅

X = df.drop(columns="target")           # id ham qoldi            # ⚠️
X = df.drop(columns=["target", "mijoz_id"])                       # ✅

permutation_importance(model, Xte, yte, n_repeats=2)              # ⚠️
permutation_importance(model, Xte, yte, n_repeats=20)             # ✅

# butun ma'lumotda muhimlik -> belgi tanlash -> CV                # ⚠️
# belgi tanlashni Pipeline ichiga qo'ying 12.9-bob                  # ✅

1.  # MDI to'liq nomi?
2.  # MDI ning birinchi tuzog'i?
3.  # ikkinchisi?
4.  # uchinchisi?
5.  # permutation qanday ishlaydi?
6.  # qaysi to'plamda?
7.  # n_repeats odatiy?
8.  # drop-column kamchiligi?
9.  # SHAP nimadan kelib chiqqan?
10. # korrelyatsiyada yechim?
11. # muhimlik sababiylikmi?
12. # ID ni nima qilish kerak?

1.  plt.barh(nomlar, model.feature_importances_)

2.  permutation_importance(model, Xtr, ytr)

3.  permutation_importance(model, Xte, yte, n_repeats=2)

4.  X = df.drop(columns="target")   # mijoz_id qoldi

5.  # "eng muhim belgini o'zgartirsak natija yaxshilanadi"