SHAP: modelni tushuntirish

import pandas as pd
import numpy as np
show(pd.DataFrame([
    {"Savol": "Model UMUMAN nimaga qaraydi?", "Nomi": "Global tushuntirish",
     "Asboblar": "feature_importances_, permutation importance, SHAP summary"},
    {"Savol": "BU QATOR nega shunday baholandi?", "Nomi": "Lokal tushuntirish",
     "Asboblar": "SHAP waterfall, LIME"},
]))

from sklearn.model_selection import train_test_split
from sklearn.ensemble import RandomForestClassifier

df = pd.read_csv(DATA / "uy_ijara.csv")
son = ["xonalar", "maydon_m2", "qavat", "qavatlar_soni", "metro_km", "mebel"]
X, y = df[son], df["ijaraga_berildi_30k"]
X_tr, X_te, y_tr, y_te = train_test_split(X, y, test_size=0.3, random_state=42, stratify=y)

model = RandomForestClassifier(n_estimators=200, min_samples_leaf=10, random_state=42).fit(X_tr, y_tr)
print(f"test accuracy: {model.score(X_te, y_te):.4f}")
print()
print("Eski usul — feature_importances_:")
for nom, v in sorted(zip(son, model.feature_importances_), key=lambda x: -x[1]):
    print(f"  {nom:16s} {v:.4f}")
print()
print("Muammo: bu bitta kvartira haqida HECH NARSA aytmaydi.")

import shap

tushuntiruvchi = shap.TreeExplainer(model)
sh = tushuntiruvchi(X_te)
qiymatlar = sh.values[:, :, 1] if sh.values.ndim == 3 else sh.values
asos = sh.base_values[0, 1] if np.ndim(sh.base_values) > 1 else sh.base_values[0]

print(f"SHAP qiymatlari shakli: {qiymatlar.shape}  -> har QATOR va har USTUN uchun bitta son")
print(f"Asos qiymat (o'rtacha bashorat): {asos:.4f}")
print()
# tenglikni tekshiramiz
i = 0
jami = asos + qiymatlar[i].sum()
haqiqiy = model.predict_proba(X_te.iloc[[i]])[0, 1]
print(f"1-qator uchun tekshiruv:")
print(f"  asos + SHAP yig'indisi = {asos:.4f} + {qiymatlar[i].sum():+.4f} = {jami:.4f}")
print(f"  modelning haqiqiy bashorati          = {haqiqiy:.4f}")
print(f"  mos keldimi: {bool(np.isclose(jami, haqiqiy, atol=1e-6))}")

i = 0
qator = X_te.iloc[i]
hissa = pd.DataFrame({
    "xususiyat": son,
    "qiymati": qator.values,
    "SHAP hissasi": qiymatlar[i].round(4),
}).sort_values("SHAP hissasi", key=abs, ascending=False)
hissa["yo'nalish"] = ["TEZ ketish tomoniga" if v > 0 else "SEKIN ketish tomoniga"
                      for v in hissa["SHAP hissasi"]]
show(hissa)
print()
print(f"O'rtacha bashorat     : {asos:.4f}")
print(f"Hissalar yig'indisi   : {qiymatlar[i].sum():+.4f}")
print(f"Bu kvartira bashorati : {model.predict_proba(X_te.iloc[[i]])[0, 1]:.4f}")
print(f"Haqiqiy javob         : {y_te.iloc[i]}")

import matplotlib.pyplot as plt
rasm, ax = plt.subplots(figsize=(9.5, 4.5))
tartib = np.argsort(np.abs(qiymatlar[i]))
ranglar = ["#137333" if qiymatlar[i][k] > 0 else "#c5221f" for k in tartib]
ax.barh(range(len(tartib)), qiymatlar[i][tartib], color=ranglar, alpha=.85)
ax.set_yticks(range(len(tartib)))
ax.set_yticklabels([f"{son[k]} = {qator.values[k]:.4g}" for k in tartib])
ax.axvline(0, color="k", lw=1)
ax.set_xlabel("SHAP hissasi (ehtimollikka ta'siri)")
ax.set_title(f"Bitta kvartira: o'rtacha {asos:.3f} -> bashorat {asos + qiymatlar[i].sum():.3f}",
             fontweight="bold")
ax.grid(alpha=.3, axis="x")
fig("shap_lokal.png", caption="Yashil — tez ketish tomoniga, qizil — sekin ketish tomoniga. Uzunlik hissaning kuchini ko'rsatadi.")

from sklearn.inspection import permutation_importance
pi = permutation_importance(model, X_te, y_te, n_repeats=20, random_state=42, scoring="f1")

taqqos = pd.DataFrame({
    "xususiyat": son,
    "feature_importances_": model.feature_importances_.round(4),
    "permutation": pi.importances_mean.round(4),
    "SHAP (|o'rtacha|)": np.abs(qiymatlar).mean(axis=0).round(4),
}).sort_values("SHAP (|o'rtacha|)", ascending=False)
show(taqqos)
print()
print("Uch ustunning tartibi mos keladimi — tekshiring.")

rasm, axes = plt.subplots(1, 2, figsize=(14, 4.8))

# summary bar
tartib_g = np.argsort(np.abs(qiymatlar).mean(axis=0))
axes[0].barh(range(len(son)), np.abs(qiymatlar).mean(axis=0)[tartib_g], color="#1a73e8", alpha=.85)
axes[0].set_yticks(range(len(son)))
axes[0].set_yticklabels([son[k] for k in tartib_g])
axes[0].set_xlabel("|SHAP| o'rtachasi")
axes[0].set_title("Global muhimlik", fontweight="bold")
axes[0].grid(alpha=.3, axis="x")

# beeswarm o'rniga: eng muhim xususiyat uchun qiymat va SHAP bog'liqligi
eng_muhim = int(np.argmax(np.abs(qiymatlar).mean(axis=0)))
axes[1].scatter(X_te.iloc[:, eng_muhim], qiymatlar[:, eng_muhim], alpha=.5, s=22, c="#c5221f")
axes[1].axhline(0, color="k", lw=1, ls="--")
axes[1].set_xlabel(son[eng_muhim])
axes[1].set_ylabel(f"SHAP hissasi")
axes[1].set_title(f"'{son[eng_muhim]}' qiymati va uning ta'siri", fontweight="bold")
axes[1].grid(alpha=.3)
fig("shap_global.png", caption="Chapda: qaysi xususiyat umuman muhim. O'ngda: eng muhim xususiyatning qiymati oshganda ta'siri qanday o'zgaradi.")

show(pd.DataFrame([
    {"Xususiyat": "Lokal (bitta qator) tushuntirish", "feature_importances_": "YO'Q", "SHAP": "BOR"},
    {"Xususiyat": "Yo'nalish (oshiradi/kamaytiradi)", "feature_importances_": "YO'Q", "SHAP": "BOR (belgi)"},
    {"Xususiyat": "Yig'indi bashoratga teng", "feature_importances_": "YO'Q", "SHAP": "BOR (kafolat)"},
    {"Xususiyat": "Ko'p qiymatli ustunga moyillik", "feature_importances_": "BOR (muammo)", "SHAP": "Kamroq"},
    {"Xususiyat": "Hisoblash narxi", "feature_importances_": "Bepul", "SHAP": "Qimmat"},
]))

import time
print("Hisoblash narxi — o'lchaymiz:")
t0 = time.perf_counter(); _ = model.feature_importances_; t1 = time.perf_counter()
print(f"  feature_importances_ : {(t1-t0)*1000:.3f} ms")
t0 = time.perf_counter(); _ = shap.TreeExplainer(model)(X_te); t1 = time.perf_counter()
print(f"  SHAP (TreeExplainer) : {(t1-t0)*1000:.1f} ms  ({len(X_te)} qator uchun)")
t0 = time.perf_counter()
_ = permutation_importance(model, X_te, y_te, n_repeats=20, random_state=42, scoring="f1")
t1 = time.perf_counter()
print(f"  permutation (20 marta): {(t1-t0)*1000:.1f} ms")

from sklearn.linear_model import LogisticRegression
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import Pipeline

sc = StandardScaler().fit(X_tr)
lr = LogisticRegression(max_iter=3000).fit(sc.transform(X_tr), y_tr)

lin_exp = shap.LinearExplainer(lr, sc.transform(X_tr))
lin_sh = lin_exp(sc.transform(X_te))

taqqos2 = pd.DataFrame({
    "xususiyat": son,
    "koeffitsiyent": lr.coef_[0].round(4),
    "SHAP |o'rtacha|": np.abs(lin_sh.values).mean(axis=0).round(4),
}).sort_values("SHAP |o'rtacha|", ascending=False)
show(taqqos2)
print()
print("Chiziqli modelda SHAP = koeffitsiyent x (qiymat - o'rtacha).")
print("Shuning uchun tartib koeffitsiyentlar mutlaq qiymati bilan MOS keladi.")

bashorat = model.predict(X_te)
xato_mask = (bashorat != y_te.to_numpy())
print(f"Xato qatorlar: {xato_mask.sum()} / {len(y_te)}")
print()
taqq = pd.DataFrame({
    "xususiyat": son,
    "to'g'ri qatorlarda |SHAP|": np.abs(qiymatlar[~xato_mask]).mean(axis=0).round(4),
    "XATO qatorlarda |SHAP|": np.abs(qiymatlar[xato_mask]).mean(axis=0).round(4),
})
taqq["farq"] = (taqq["XATO qatorlarda |SHAP|"] - taqq["to'g'ri qatorlarda |SHAP|"]).round(4)
show(taqq.sort_values("farq", key=abs, ascending=False))
print()
print("Eng katta farqli xususiyat — model aynan unda chalg'iydi.")

show(pd.DataFrame([
    {"Tushuncha": "Global vs lokal", "Bitta jumla": "Model umuman nimaga qaraydi / bu qator nega shunday"},
    {"Tushuncha": "SHAP g'oyasi", "Bitta jumla": "Bashoratni o'rtachadan farqini adolatli taqsimlash"},
    {"Tushuncha": "Kafolat", "Bitta jumla": "asos + SHAP yig'indisi = AYNAN bashorat"},
    {"Tushuncha": "Belgi", "Bitta jumla": "Musbat — target tomoniga, manfiy — teskari"},
    {"Tushuncha": "TreeExplainer", "Bitta jumla": "Daraxtlar uchun aniq va tez"},
    {"Tushuncha": "KernelExplainer", "Bitta jumla": "Har qanday model, lekin sekin — namuna oling"},
    {"Tushuncha": "Eng katta xato", "Bitta jumla": "SHAP ni SABABIYLIK deb o'qish"},
    {"Tushuncha": "Amaliy foyda", "Bitta jumla": "Xato tahlilini xususiyat darajasiga tushiradi"},
]))