Xato tahlili: natija yomon bo'lsa nima qilish

import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.pipeline import Pipeline
from sklearn.compose import ColumnTransformer
from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.impute import SimpleImputer
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import f1_score, accuracy_score

df = pd.read_csv(DATA / "uy_ijara.csv")
son = ["xonalar", "maydon_m2", "qavat", "qavatlar_soni", "metro_km", "mebel", "qurilgan_yil"]
X, y = df[son + ["tuman"]], 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)

quvur = Pipeline([
    ("tayyorla", ColumnTransformer([
        ("son", Pipeline([("toldir", SimpleImputer(strategy="median")),
                          ("shkala", StandardScaler())]), son),
        ("matn", OneHotEncoder(handle_unknown="ignore"), ["tuman"]),
    ])),
    ("model", LogisticRegression(max_iter=2000)),
]).fit(X_tr, y_tr)

bashorat = quvur.predict(X_te)
ehtimol = quvur.predict_proba(X_te)[:, 1]
print(f"Accuracy: {accuracy_score(y_te, bashorat):.4f}")
print(f"F1      : {f1_score(y_te, bashorat):.4f}")
print()
print("Endi: BU RAQAMLARGA EMAS, XATOLARGA qaraymiz.")

tahlil = X_te.copy()
tahlil["haqiqat"] = y_te
tahlil["bashorat"] = bashorat
tahlil["ishonch"] = ehtimol.round(3)
tahlil["xato"] = (tahlil.haqiqat != tahlil.bashorat).astype(int)

print(f"Jami xato: {tahlil.xato.sum()} / {len(tahlil)} ({tahlil.xato.mean():.1%})")
print()
print("Model ENG ISHONCH bilan adashgan 6 qator:")
eng = tahlil[tahlil.xato == 1].assign(
    xatolik=lambda d: (d.ishonch - d.haqiqat).abs()).nlargest(6, "xatolik")
show(eng[["tuman", "xonalar", "maydon_m2", "metro_km", "mebel", "haqiqat", "bashorat", "ishonch"]])

def kesib_kor(ustun, nomi=None):
    g = tahlil.groupby(ustun, observed=True).agg(jami=("xato", "size"), xato=("xato", "sum"))
    g["xato_ulushi"] = (g.xato / g.jami).round(3)
    g = g[g.jami >= 10].sort_values("xato_ulushi", ascending=False)
    print(f"--- {nomi or ustun} bo'yicha ---")
    print(g.to_string())
    print()

kesib_kor("tuman")
kesib_kor("xonalar")

tahlil["maydon_guruh"] = pd.cut(tahlil.maydon_m2, bins=[0, 40, 60, 80, 300],
                                labels=["<40", "40-60", "60-80", "80+"])
tahlil["metro_guruh"] = pd.cut(tahlil.metro_km, bins=[0, 1, 2.5, 5, 20],
                               labels=["<1km", "1-2.5km", "2.5-5km", "5km+"])
kesib_kor("maydon_guruh", "Maydon")
kesib_kor("metro_guruh", "Metroga masofa")

fp = tahlil[(tahlil.haqiqat == 0) & (tahlil.bashorat == 1)]
fn = tahlil[(tahlil.haqiqat == 1) & (tahlil.bashorat == 0)]

print(f"FP (yo'qni bor dedi): {len(fp)} ta")
print(f"FN (borni yo'q dedi): {len(fn)} ta")
print()
qatorlar = []
for u in ["maydon_m2", "metro_km", "xonalar", "mebel", "qavat"]:
    qatorlar.append({"Ustun": u,
                     "FP o'rtacha": round(fp[u].mean(), 2),
                     "FN o'rtacha": round(fn[u].mean(), 2),
                     "To'g'ri o'rtacha": round(tahlil[tahlil.xato == 0][u].mean(), 2)})
show(pd.DataFrame(qatorlar))

import matplotlib.pyplot as plt

rasm, axes = plt.subplots(1, 2, figsize=(13, 4.5))

axes[0].hist([tahlil[tahlil.haqiqat == 0].ishonch, tahlil[tahlil.haqiqat == 1].ishonch],
             bins=20, label=["haqiqatda sekin", "haqiqatda tez"], color=["#c5221f", "#137333"],
             alpha=.75, stacked=False)
axes[0].axvline(0.5, color="k", ls="--", label="chegara 0.5")
axes[0].set_xlabel("Model ishonchi"); axes[0].set_ylabel("Qatorlar soni")
axes[0].set_title("Ishonch taqsimoti"); axes[0].legend(); axes[0].grid(alpha=.3)

guruh = pd.cut(tahlil.ishonch, bins=np.arange(0, 1.01, 0.1))
xato_ulush = tahlil.groupby(guruh, observed=True).xato.mean()
axes[1].bar(range(len(xato_ulush)), xato_ulush.values, color="#1a73e8", alpha=.8)
axes[1].set_xticks(range(len(xato_ulush)))
axes[1].set_xticklabels([f"{i/10:.1f}" for i in range(len(xato_ulush))], rotation=45)
axes[1].set_xlabel("Ishonch oralig'i"); axes[1].set_ylabel("Xato ulushi")
axes[1].set_title("Qaysi ishonchda ko'p adashadi"); axes[1].grid(alpha=.3, axis="y")

fig("xato_tahlili.png", caption="Chapda: ikki sinf ishonch bo'yicha qanchalik ajralgan. O'ngda: xato qaysi ishonch oralig'ida to'plangan.")

from sklearn.model_selection import learning_curve

hajmlar, train_ball, val_ball = learning_curve(
    quvur, X, y, train_sizes=np.linspace(0.15, 1.0, 8), cv=5,
    scoring="f1", shuffle=True, random_state=42)

jadval = pd.DataFrame({
    "train hajmi": hajmlar,
    "train F1": train_ball.mean(axis=1).round(4),
    "val F1": val_ball.mean(axis=1).round(4),
    "farq (GAP)": (train_ball.mean(axis=1) - val_ball.mean(axis=1)).round(4),
})
show(jadval)

rasm2, ax = plt.subplots(figsize=(9, 5))
ax.plot(hajmlar, train_ball.mean(axis=1), "b-o", lw=2, label="Train F1")
ax.plot(hajmlar, val_ball.mean(axis=1), "r-o", lw=2, label="Validation F1")
ax.fill_between(hajmlar, train_ball.mean(axis=1), val_ball.mean(axis=1), alpha=.15, color="red")
ax.set_xlabel("Train qatorlari soni"); ax.set_ylabel("F1")
ax.set_title("Learning curve — ko'proq ma'lumot yordam beradimi?")
ax.legend(); ax.grid(alpha=.3)
fig("learning_curve.png", caption="Ikki chiziq yaqinlashib tekislangan bo'lsa, ko'proq ma'lumot yordam bermaydi.")

show(pd.DataFrame([
    {"Ko'rinish": "Ikki chiziq YAQIN va ikkisi ham PAST", "Diagnoz": "Underfitting (high bias)",
     "Harakat": "Kuchliroq model, ko'proq xususiyat. Ko'proq MA'LUMOT YORDAM BERMAYDI"},
    {"Ko'rinish": "Train BALAND, validation PAST, oraliq KATTA", "Diagnoz": "Overfitting (high variance)",
     "Harakat": "Ko'proq ma'lumot, regularizatsiya, soddaroq model"},
    {"Ko'rinish": "Validation hali O'SIB boryapti", "Diagnoz": "Ma'lumot kam",
     "Harakat": "Ko'proq ma'lumot yig'ish FOYDALI"},
    {"Ko'rinish": "Validation TEKISLANGAN", "Diagnoz": "Ma'lumot yetarli",
     "Harakat": "Yangi ustun yoki boshqa model kerak"},
]))

oxirgi_gap = train_ball.mean(axis=1)[-1] - val_ball.mean(axis=1)[-1]
oxirgi_ikki = val_ball.mean(axis=1)[-1] - val_ball.mean(axis=1)[-3]
print(f"Oxirgi nuqtada GAP            : {oxirgi_gap:.4f}")
print(f"Validation oxirgi o'sishi     : {oxirgi_ikki:+.4f}")
print()
if oxirgi_gap > 0.08:
    print("DIAGNOZ: overfitting belgisi -> ko'proq ma'lumot yoki regularizatsiya")
elif oxirgi_ikki > 0.01:
    print("DIAGNOZ: validation hali o'syapti -> ko'proq ma'lumot FOYDALI")
else:
    print("DIAGNOZ: ma'lumot yetarli, chiziqlar yaqin -> YANGI USTUN yoki boshqa model kerak")

show(pd.DataFrame([
    {"Dalil": "Xatolar bitta guruhda to'plangan", "Harakat": "O'sha guruh uchun yangi ustun yoki alohida model"},
    {"Dalil": "FP va FN o'rtachalari farq qiladi", "Harakat": "Threshold ni surish (8-bob)"},
    {"Dalil": "Model ishonch bilan adashadi", "Harakat": "Yorliqlarni tekshirish, muhim ustunni qo'shish"},
    {"Dalil": "Learning curve: validation o'syapti", "Harakat": "Ko'proq ma'lumot yig'ish"},
    {"Dalil": "Learning curve: GAP katta", "Harakat": "Regularizatsiya yoki soddaroq model"},
    {"Dalil": "Learning curve: ikkisi ham past va tekis", "Harakat": "Kuchliroq model (daraxt, boosting)"},
    {"Dalil": "Dalil yo'q, hammasi normal", "Harakat": "Tugating. Model yetarli (39-bob)"},
]))

from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import cross_val_score

def quvur_bilan(model):
    return Pipeline([
        ("tayyorla", ColumnTransformer([
            ("son", Pipeline([("toldir", SimpleImputer(strategy="median")),
                              ("shkala", StandardScaler())]), son),
            ("matn", OneHotEncoder(handle_unknown="ignore"), ["tuman"]),
        ])),
        ("model", model),
    ])

lr = cross_val_score(quvur_bilan(LogisticRegression(max_iter=2000)), X, y, cv=5, scoring="f1")
rf = cross_val_score(quvur_bilan(RandomForestClassifier(n_estimators=200, random_state=42)),
                     X, y, cv=5, scoring="f1")
print(f"Logistic Regression: {lr.mean():.4f} ±{lr.std():.4f}")
print(f"Random Forest      : {rf.mean():.4f} ±{rf.std():.4f}")
print(f"Farq               : {rf.mean() - lr.mean():+.4f}  (tarqalish ±{max(lr.std(), rf.std()):.4f})")
print()
print("Hukm:", "Random Forest ishonchli yaxshiroq" if rf.mean() - lr.mean() > max(lr.std(), rf.std())
      else "Farq tarqalish ichida — oddiyroq modelni saqlang")

daftar = """
+----------------------------------------------------------------------+
|  XATO TAHLILI DAFTARI                                                |
+----------------------------------------------------------------------+
|  Sana        :                                                       |
|  Model       :                                                       |
|  Metrika     : F1 = ______  (CV: ______ +- ______)                   |
|  Baza        : ______                                                |
+----------------------------------------------------------------------+
|  1. Xatolar qaysi guruhda to'plangan?                                |
|     ................................................................ |
|  2. FP va FN farqi nimada?                                           |
|     ................................................................ |
|  3. Model ishonch bilan adashadimi? (0.9+ da xato bormi)             |
|     ................................................................ |
|  4. Learning curve diagnozi?                                         |
|     [ ] underfitting  [ ] overfitting  [ ] ma'lumot kam  [ ] yetarli |
+----------------------------------------------------------------------+
|  QAROR (faqat BITTASI):                                              |
|     [ ] yangi xususiyat: ...........................................  |
|     [ ] boshqa model: ..............................................  |
|     [ ] ko'proq ma'lumot                                             |
|     [ ] threshold surish: ..........................................  |
|     [ ] tugatish                                                     |
+----------------------------------------------------------------------+
|  Natija: ______  (oldingi: ______)   Farq: ______                    |
|  Gipotezam to'g'ri chiqdimi?  [ ] ha   [ ] YO'Q <- eng qimmatli      |
+----------------------------------------------------------------------+
"""
print(daftar)

show(pd.DataFrame([
    {"Qadam": "1. Xatolarni o'qi", "Vosita": "nlargest(ishonch) — ko'zingiz bilan"},
    {"Qadam": "2. Guruhlab sana", "Vosita": "groupby(ustun).xato.mean()"},
    {"Qadam": "3. FP va FN ni ajrat", "Vosita": "Ikki xil xato, ikki xil sabab"},
    {"Qadam": "4. Ishonch taqsimoti", "Vosita": "Model qayerda shubhada, qayerda ishonch bilan adashadi"},
    {"Qadam": "5. Learning curve", "Vosita": "Ko'proq ma'lumot kerakmi — TAXMIN emas, O'LCHOV"},
    {"Qadam": "6. BITTA o'zgarish", "Vosita": "Dalilga asoslangan qaror, keyin 9-qadamga qaytish"},
]))