23.5-dars: Matn klassifikatsiyasi — klassik bazaviylar

MATN KLASSIFIKATSIYASI:
  kirish:  matn (turli uzunlikdagi so'zlar ketma-ketligi)
  chiqish: sinf (salbiy / neytral / ijobiy, spam / spam emas, mavzu ...)

ZANJIR:
  xom matn -> normallashtirish 23.1-bob -> tokenlar 23.2-bob
           -> vektor (TF-IDF, 23.4) -> klassifikator -> sinf

ENG ODDIY BAZAVIY (DummyClassifier):
  hamma sharhga eng ko'p uchraydigan sinfni aytadi
  3 ta teng sinfda aniqlik ~ 0.33
  har qanday model shundan ANCHA yaxshi bo'lishi kerak

BIZNING KORPUS:
  3000 sharh, 3 sinf (taxminan teng)
  yorliq = gaplar "ball"larining o'rtachasi (+1, 0, -1) + ozgina shovqin
    o'rtacha > 0.35 -> ijobiy, < -0.35 -> salbiy, aks holda neytral
    "... sifatli ham ... sifatli boshqa olmayman" -> (1 + 1 - 1) / 3 = 0.33
    -> NEYTRAL (ikki maqtov va bitta shikoyat)
  inkor: "sifatli emas" -> salbiy, "sust emas" -> ijobiy
  imlo xatolari: so'zlarning ~7% ida (tushib qolgan, ikkilangan,
                 o'rin almashgan harf)

TF-IDF HAM O'RGANADI:
  fit:  lug'at (qaysi so'zlar/n-gramlar bor) + IDF (har so'z nechta
        hujjatda uchraydi)
  bu statistikalar - modelning bir qismi

NOTO'G'RI (leakage, 19-qism):
  vek = TfidfVectorizer().fit(barcha_matnlar)   # test ham ichida
  X = vek.transform(barcha_matnlar)
  train_test_split(X, y) ...
  -> lug'atda FAQAT testda uchragan so'zlar bor
  -> IDF test hujjatlarini ham sanagan

TO'G'RI:
  model = Pipeline([("vek", TfidfVectorizer()),
                    ("clf", LogisticRegression())])
  model.fit(X_oquv, y_oquv)        # vektorizator faqat o'quvda
  cross_val_score(model, X, y)     # har foldda QAYTA fit

AMALDA:
  TF-IDF dagi leakage odatda kichik natija farqini beradi
  lekin u "kichik" ekanini bilmaysiz - tekshirmaguncha
  Pipeline bilan bu savol umuman tug'ilmaydi

LOGISTIK REGRESSIYA (14-qism):
  har sinf uchun: ball_k = w_k . x + b_k,  softmax -> ehtimollar
  C - regularizatsiya teskarisi (katta C = kuchsiz regularizatsiya)
  + ehtimol beradi, koeffitsiyentlarni o'qish mumkin 2.6-bob

MULTINOMIAL NAIVE BAYES:
  P(sinf | matn) ~ P(sinf) * ko'paytma P(so'z | sinf)
  "so'zlar sinf ichida bir-biriga bog'liq emas" degan sodda faraz
  alpha - silliqlash (ko'rilmagan so'z uchun nol ehtimol bo'lmasin)
  + juda tez, juda kam ma'lumotda ham ishlaydi
  - "sifatli" va "emas" ni alohida sanaydi - inkorni tushunmaydi

LinearSVC:
  sinflar orasidagi chegarani maksimal "oraliq" bilan qidiradi
  C - regularizatsiya teskarisi
  + siyrak, ko'p o'lchamli matn vektorlarida kuchli
  - ehtimol bermaydi (faqat decision_function)

UMUMIY XOSSA:
  uchalasi ham CHIZIQLI: har belgi (so'z/n-gram) o'z og'irligini oladi
  matn vektori 10 000+ o'lchamli bo'lsa ham o'rgatish soniyalar oladi

SO'Z UNIGRAMLARI (ngram_range=(1, 1)):
  "sifati sifatli emas" -> {sifati, sifatli, emas}
  "sifatli" ijobiy og'irlik oladi, "emas" - deyarli neytral
  -> inkor YO'QOLADI

SO'Z BIGRAMLARI (ngram_range=(1, 2)):
  + {sifati sifatli, sifatli emas}
  "sifatli emas" - alohida belgi, o'z (salbiy) og'irligi bilan

BELGI N-GRAMLARI:
  analyzer="char_wb", (2, 4): faqat SO'Z ICHIDA
    "sifatli" -> " s", "si", "if", ... "atli ", ...
    so'zlar chegarasidan o'tmaydi -> "emas" bilan bog'lanish yo'q
  analyzer="char", (2, 4): bo'shliqdan ham o'tadi
    "li em", "i ema" -> inkor ham ko'rinadi

IMLO XATOSI:
  "sifatli" -> "sifatil" (harflar o'rin almashgan)
  so'z darajasida: butunlay yangi so'z, lug'atda yo'q -> e'tiborsiz
  belgi darajasida: "sif", "ifa", "fat" saqlanadi -> signal qisman qoladi

NARX:
  har matnda nol bo'lmagan belgilar ~10 barobar ko'p
  (bizning korpusda o'rtacha: so'z+bigram ~24, belgi (2, 4) ~230)
  -> vektorlash, o'rgatish va bashorat sekinroq

BITTA BO'LINISH YETMAYDI:
  1000 ta test sharhida aniqlikning SE si ~ sqrt(0.8*0.2/1000) ~ 0.013
  0.01 farq - shovqin ichida bo'lishi mumkin

JUFTLASHGAN CV (18-qism):
  cv = StratifiedKFold(5, shuffle=True, random_state=0)
  HAMMA modelga BIR XIL foldlar
  har foldda: d_k = aniq_A_k - aniq_B_k
  o'rtacha d, SE = std(d, ddof=1) / sqrt(5)
  |d| > 2 * SE -> sezilarli (taxminiy qoida)

QAROR QOIDASI:
  1. eng yaxshi o'rtacha natijali modelni toping
  2. har modelni u bilan juftlab solishtiring
  3. "sezilarli yomon EMAS" modellar ichidan ENG SODDASINI tanlang

SODDALIK TARTIBI (bizning darsda):
  NB so'z < LR so'z < NB so'z+bigram < LR so'z+bigram < SVC so'z+bigram
  < LR belgi (so'z ichida) < LR belgi
  (belgi modellari: lug'at katta, sekin, tushuntirish qiyin)

KOEFFITSIYENTLAR (logistik regressiya):
  clf.coef_ shakli (n_sinf, n_belgi)
  har sinf uchun eng katta musbat og'irlikli belgilar -
  "bu sinfga eng kuchli ovoz beradigan so'zlar"
  vek.get_feature_names_out() - indeks -> so'z

NIMANI TEKSHIRAMIZ:
  mantiqiy so'zlarmi? ("zo'r", "sust emas" -> ijobiy)
  "shubhali" so'zlarmi? (mahsulot nomi, sana, do'kon nomi ->
    ma'lumotda noxolis bog'liqlik yoki leakage belgisi)

XATO TAHLILI:
  chalkashlik matritsasi: qaysi sinflar adashadi?
    odatda neytral <-> ijobiy/salbiy, salbiy <-> ijobiy kam
  guruhlar bo'yicha aniqlik: inkorli / inkorsiz, qisqa / uzun
  eng ISHONCHLI xatolar: ko'pincha yorliq xatosi yoki model
    tushunmaydigan tuzilish ("... emas", "... lekin ...")

KEYINGI QADAM:
  xato guruhi -> belgi (bigram), ma'lumot yoki boshqa model

from sklearn.dummy import DummyClassifier
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import StratifiedKFold, cross_val_score
from sklearn.naive_bayes import MultinomialNB
from sklearn.pipeline import make_pipeline
from sklearn.svm import LinearSVC

TP = r"[\w']+"                     # apostrofli so'zlar butun qoladi

soz = make_pipeline(
    TfidfVectorizer(token_pattern=TP, ngram_range=(1, 2)),
    LogisticRegression(C=10, max_iter=2000))
belgi = make_pipeline(
    TfidfVectorizer(analyzer="char", ngram_range=(2, 4)),
    LogisticRegression(C=10, max_iter=2000))
nb = make_pipeline(TfidfVectorizer(token_pattern=TP, ngram_range=(1, 2)),
                   MultinomialNB(alpha=0.1))
svc = make_pipeline(TfidfVectorizer(token_pattern=TP, ngram_range=(1, 2)),
                    LinearSVC(C=0.5))

cv = StratifiedKFold(5, shuffle=True, random_state=0)
a = cross_val_score(soz, matnlar, y, cv=cv)      # bir xil foldlar
b = cross_val_score(belgi, matnlar, y, cv=cv)
d = b - a
se = d.std(ddof=1) / len(d) ** 0.5

vek, clf = soz.fit(X_oquv, y_oquv).named_steps.values()
nomlar = vek.get_feature_names_out()
top = nomlar[clf.coef_[2].argsort()[::-1][:10]]  # ijobiy sinf so'zlari

DummyClassifier -> TF-IDF + LogReg (Pipeline) -> n-gramlar
token_pattern apostrofni saqlasin
bigram - inkor, belgi n-gram - imlo xatosi
juftlashgan CV: bir xil foldlar, farq + SE
sezilarli yomon bo'lmagan ENG SODDA model
koeffitsiyentlar va eng ishonchli xatolarni ko'ring

"""Sintetik o'zbekcha sharhlar: bazaviy, Pipeline va leakage."""

import numpy as np
from sklearn.dummy import DummyClassifier
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import accuracy_score, f1_score
from sklearn.model_selection import train_test_split
from sklearn.pipeline import make_pipeline

IJOBIY = ["yaxshi", "zo'r", "a'lo", "ajoyib", "sifatli", "qulay",
          "chiroyli", "mustahkam"]
SALBIY = ["yomon", "sifatsiz", "noqulay", "xunuk", "nosoz", "mo'rt",
          "sust", "chatoq"]
NEYTRAL = ["o'rtacha", "oddiy", "normal", "odatiy", "kutilgandek"]
JIHAT = ["sifati", "narxi", "ekrani", "batareyasi", "ovozi", "qadoqlash",
         "yetkazib berish", "sotuvchi", "dizayni", "materiali"]
MAHSULOT = ["telefon", "noutbuk", "quloqchin", "soat", "choynak", "kurtka",
            "krossovka", "sumka", "planshet", "kolonka"]
BOSH = ["kecha", "o'tgan hafta", "bir oy oldin", "do'kondan", "internetdan",
        "sovg'a uchun", "akam uchun", ""]
FEL = ["oldim", "buyurtma qildim", "sotib oldim", "ishlatib ko'rdim"]
YAKUN = {1: ["tavsiya qilaman", "yana olaman", "hammaga maslahat beraman"],
         -1: ["tavsiya qilmayman", "boshqa olmayman", "pulimga achinaman"],
         0: ["hozircha shu", "vaqt ko'rsatadi", "yana ishlatib ko'ramiz"]}
ULUSH = {1: [0.8, 0.1, 0.1], -1: [0.1, 0.8, 0.1], 0: [0.15, 0.15, 0.7]}
SINFLAR = ["salbiy", "neytral", "ijobiy"]
TP = r"[\w']+"


def gap(qutb, rng):
    j = rng.choice(JIHAT)
    juda = "juda " if rng.random() < 0.3 else ""
    r = rng.random()
    if qutb == 1:
        if r < 0.2:
            return f"{j} {rng.choice(SALBIY)} emas"
        if r < 0.3:
            return f"{j} bilan muammo yo'q"
        return f"{j} {juda}{rng.choice(IJOBIY)}"
    if qutb == -1:
        if r < 0.25:
            return f"{j} {rng.choice(IJOBIY)} emas"
        if r < 0.35:
            return f"{j} bilan muammo bor"
        return f"{j} {juda}{rng.choice(SALBIY)}"
    return f"{j} {rng.choice(NEYTRAL)}"


def sharh(rng):
    """Kayfiyat -> gaplar -> yorliq (gaplar ballining o'rtachasi)."""
    kayfiyat = int(rng.integers(-1, 2))
    qismlar = [f"{rng.choice(BOSH)} {rng.choice(MAHSULOT)} "
               f"{rng.choice(FEL)}".strip()]
    ballar = []
    for t in range(int(rng.integers(1, 5))):
        q = int(rng.choice([1, -1, 0], p=ULUSH[kayfiyat]))
        bog = rng.choice(["lekin", "va", "ammo", "ham", ""]) if t else ""
        qismlar.append(f"{bog} {gap(q, rng)}".strip())
        ballar.append(q)
    if rng.random() < 0.5:
        q = int(rng.choice([1, -1, 0], p=ULUSH[kayfiyat]))
        qismlar.append(rng.choice(YAKUN[q]))
        ballar.append(q)
    ball = np.mean(ballar) + rng.normal(0, 0.1)
    return " ".join(qismlar), 2 if ball > 0.35 else (0 if ball < -0.35 else 1)


def imlo_xato(matn, ulush, rng):
    chiqish = []
    for s in matn.split():
        if len(s) >= 4 and rng.random() < ulush:
            i = int(rng.integers(1, len(s) - 1))
            tur = int(rng.integers(0, 3))
            if tur == 0:
                s = s[:i] + s[i + 1:]
            elif tur == 1:
                s = s[:i] + s[i] + s[i:]
            else:
                s = s[:i - 1] + s[i] + s[i - 1] + s[i + 1:]
        chiqish.append(s)
    return " ".join(chiqish)


def korpus(n, seed=0, xato=0.07):
    rng = np.random.default_rng(seed)
    toza, y = zip(*[sharh(rng) for _ in range(n)])
    rng2 = np.random.default_rng(seed + 1000)
    return [imlo_xato(t, xato, rng2) for t in toza], np.array(y), list(toza)


def main() -> None:
    matnlar, y, toza = korpus(3000)

    print("=== 1. Korpus ===")
    soni = np.bincount(y)
    for k, nom in enumerate(SINFLAR):
        print(f"  {nom:<8} {soni[k]:>5} ({soni[k] / len(y):.1%})")
    uzun = np.array([len(m.split()) for m in matnlar])
    print(f"  so'zlar soni: min {uzun.min()}, mediana "
          f"{int(np.median(uzun))}, max {uzun.max()}")
    toza_soz = [s for t in toza for s in t.split()]
    xato_soz = [s for t in matnlar for s in t.split()]
    ozgargan = np.mean([a != b for a, b in zip(toza_soz, xato_soz)])
    print(f"  imlo xatosi bor so'zlar: {ozgargan:.1%}")
    for k in (2, 1, 0):
        i = int(np.where(y == k)[0][0])
        print(f"  [{SINFLAR[k]}] {matnlar[i]}")

    Xtr, Xte, ytr, yte = train_test_split(matnlar, y, test_size=0.2,
                                          random_state=0, stratify=y)
    print(f"\n  o'quv {len(Xtr)}, test {len(Xte)}")

    print("\n=== 2. Eng oddiy bazaviy ===")
    dummy = DummyClassifier(strategy="most_frequent").fit(Xtr, ytr)
    print(f"  DummyClassifier aniqligi: "
          f"{accuracy_score(yte, dummy.predict(Xte)):.4f}")

    print("\n=== 3. Pipeline: TF-IDF (so'z + bigram) + LogReg ===")
    model = make_pipeline(
        TfidfVectorizer(token_pattern=TP, ngram_range=(1, 2)),
        LogisticRegression(C=10, max_iter=2000))
    model.fit(Xtr, ytr)
    p = model.predict(Xte)
    print(f"  test aniqligi: {accuracy_score(yte, p):.4f}")
    f1 = f1_score(yte, p, average=None)
    for k, nom in enumerate(SINFLAR):
        print(f"  F1 {nom:<8} {f1[k]:.4f}")
    print(f"  lug'at hajmi (faqat o'quv): "
          f"{len(model[0].vocabulary_)} belgi")

    print("\n=== 4. Sukut token_pattern va apostrof ===")
    sukut = TfidfVectorizer().fit(["sotuvchi zo'r va a'lo"])
    print(f"  sukut:   {sorted(sukut.vocabulary_)}")
    bizniki = TfidfVectorizer(token_pattern=TP).fit(["sotuvchi zo'r va a'lo"])
    print(f"  [\\w']+: {sorted(bizniki.vocabulary_)}")

    print("\n=== 5. Leakage: vektorizator butun ma'lumotda ===")
    vek_xato = TfidfVectorizer(token_pattern=TP, ngram_range=(1, 2))
    vek_xato.fit(matnlar)
    faqat_test = set(vek_xato.vocabulary_) - set(model[0].vocabulary_)
    print(f"  butun ma'lumot lug'ati: {len(vek_xato.vocabulary_)} belgi")
    print(f"  shundan faqat testda uchraganlari: {len(faqat_test)}")
    misollar = sorted(f for f in faqat_test if " " not in f)[:5]
    print(f"  masalan: {misollar}")
    clf = LogisticRegression(C=10, max_iter=2000)
    clf.fit(vek_xato.transform(Xtr), ytr)
    aniq_xato = accuracy_score(yte, clf.predict(vek_xato.transform(Xte)))
    aniq_togri = accuracy_score(yte, p)
    print(f"  leakage bilan test aniqligi: {aniq_xato:.4f} "
          f"(Pipeline: {aniq_togri:.4f})")
    if abs(aniq_xato - aniq_togri) < 0.01:
        print("  bu yerda farq 0.01 dan kichik - lekin buni faqat "
              "o'lchab bildik")
    else:
        print("  farq sezilarli - leakage bahoni buzdi")
    print("  ⭐ Vektorizator Pipeline ichida - leakage savoli tug'ilmaydi")


if __name__ == "__main__":
    main()

=== 1. Korpus ===
  salbiy     941 (31.4%)
  neytral   1040 (34.7%)
  ijobiy    1019 (34.0%)
  so'zlar soni: min 4, mediana 13, max 25
  imlo xatosi bor so'zlar: 6.6%
  [ijobiy] sumka sotib oldim qadoqlash xunuk emas
  [neytral] sovg'a uchun kurtka buyurtma qildim sifati sifatli ham yetkazib berish sifatli boshqa olmayman
  [salbiy] akam uchun noutbuk oldim sifati a'lo emas lekin yetkazib berish sust va dizayni kutilganek ham batareyasi juda sust boshqa olmayman

  o'quv 2400, test 600

=== 2. Eng oddiy bazaviy ===
  DummyClassifier aniqligi: 0.3467

=== 3. Pipeline: TF-IDF (so'z + bigram) + LogReg ===
  test aniqligi: 0.8150
  F1 salbiy   0.8248
  F1 neytral  0.7536
  F1 ijobiy   0.8675
  lug'at hajmi (faqat o'quv): 4612 belgi

=== 4. Sukut token_pattern va apostrof ===
  sukut:   ['lo', 'sotuvchi', 'va', 'zo']
  [\w']+: ["a'lo", 'sotuvchi', 'va', "zo'r"]

=== 5. Leakage: vektorizator butun ma'lumotda ===
  butun ma'lumot lug'ati: 5251 belgi
  shundan faqat testda uchraganlari: 639
  masalan: ['aakm', 'achinamaan', 'ahinaman', 'amteriali', 'batareasi']
  leakage bilan test aniqligi: 0.8167 (Pipeline: 0.8150)
  bu yerda farq 0.01 dan kichik - lekin buni faqat o'lchab bildik
  ⭐ Vektorizator Pipeline ichida - leakage savoli tug'ilmaydi

"""8 ta klassik model, bir xil 5 fold, juftlashgan farq va qaror."""

import warnings

import numpy as np
from sklearn.exceptions import ConvergenceWarning
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import StratifiedKFold, cross_validate
from sklearn.naive_bayes import MultinomialNB
from sklearn.pipeline import make_pipeline
from sklearn.svm import LinearSVC

IJOBIY = ["yaxshi", "zo'r", "a'lo", "ajoyib", "sifatli", "qulay",
          "chiroyli", "mustahkam"]
SALBIY = ["yomon", "sifatsiz", "noqulay", "xunuk", "nosoz", "mo'rt",
          "sust", "chatoq"]
NEYTRAL = ["o'rtacha", "oddiy", "normal", "odatiy", "kutilgandek"]
JIHAT = ["sifati", "narxi", "ekrani", "batareyasi", "ovozi", "qadoqlash",
         "yetkazib berish", "sotuvchi", "dizayni", "materiali"]
MAHSULOT = ["telefon", "noutbuk", "quloqchin", "soat", "choynak", "kurtka",
            "krossovka", "sumka", "planshet", "kolonka"]
BOSH = ["kecha", "o'tgan hafta", "bir oy oldin", "do'kondan", "internetdan",
        "sovg'a uchun", "akam uchun", ""]
FEL = ["oldim", "buyurtma qildim", "sotib oldim", "ishlatib ko'rdim"]
YAKUN = {1: ["tavsiya qilaman", "yana olaman", "hammaga maslahat beraman"],
         -1: ["tavsiya qilmayman", "boshqa olmayman", "pulimga achinaman"],
         0: ["hozircha shu", "vaqt ko'rsatadi", "yana ishlatib ko'ramiz"]}
ULUSH = {1: [0.8, 0.1, 0.1], -1: [0.1, 0.8, 0.1], 0: [0.15, 0.15, 0.7]}
TP = r"[\w']+"


def gap(qutb, rng):
    j = rng.choice(JIHAT)
    juda = "juda " if rng.random() < 0.3 else ""
    r = rng.random()
    if qutb == 1:
        if r < 0.2:
            return f"{j} {rng.choice(SALBIY)} emas"
        if r < 0.3:
            return f"{j} bilan muammo yo'q"
        return f"{j} {juda}{rng.choice(IJOBIY)}"
    if qutb == -1:
        if r < 0.25:
            return f"{j} {rng.choice(IJOBIY)} emas"
        if r < 0.35:
            return f"{j} bilan muammo bor"
        return f"{j} {juda}{rng.choice(SALBIY)}"
    return f"{j} {rng.choice(NEYTRAL)}"


def sharh(rng):
    kayfiyat = int(rng.integers(-1, 2))
    qismlar = [f"{rng.choice(BOSH)} {rng.choice(MAHSULOT)} "
               f"{rng.choice(FEL)}".strip()]
    ballar = []
    for t in range(int(rng.integers(1, 5))):
        q = int(rng.choice([1, -1, 0], p=ULUSH[kayfiyat]))
        bog = rng.choice(["lekin", "va", "ammo", "ham", ""]) if t else ""
        qismlar.append(f"{bog} {gap(q, rng)}".strip())
        ballar.append(q)
    if rng.random() < 0.5:
        q = int(rng.choice([1, -1, 0], p=ULUSH[kayfiyat]))
        qismlar.append(rng.choice(YAKUN[q]))
        ballar.append(q)
    ball = np.mean(ballar) + rng.normal(0, 0.1)
    return " ".join(qismlar), 2 if ball > 0.35 else (0 if ball < -0.35 else 1)


def imlo_xato(matn, ulush, rng):
    chiqish = []
    for s in matn.split():
        if len(s) >= 4 and rng.random() < ulush:
            i = int(rng.integers(1, len(s) - 1))
            tur = int(rng.integers(0, 3))
            if tur == 0:
                s = s[:i] + s[i + 1:]
            elif tur == 1:
                s = s[:i] + s[i] + s[i:]
            else:
                s = s[:i - 1] + s[i] + s[i - 1] + s[i + 1:]
        chiqish.append(s)
    return " ".join(chiqish)


def korpus(n, seed=0, xato=0.07):
    rng = np.random.default_rng(seed)
    toza, y = zip(*[sharh(rng) for _ in range(n)])
    rng2 = np.random.default_rng(seed + 1000)
    return [imlo_xato(t, xato, rng2) for t in toza], np.array(y), list(toza)


def soz(ng):
    return TfidfVectorizer(token_pattern=TP, ngram_range=ng)


def belgi(analyzer):
    return TfidfVectorizer(analyzer=analyzer, ngram_range=(2, 4))


def lr():
    return LogisticRegression(C=10, max_iter=3000)


def main() -> None:
    matnlar, y, _ = korpus(3000)
    # modellar SODDALIK tartibida (2.5-bo'lim)
    modellar = {
        "NB so'z": make_pipeline(soz((1, 1)), MultinomialNB(alpha=0.1)),
        "LR so'z": make_pipeline(soz((1, 1)), lr()),
        "NB so'z+bi": make_pipeline(soz((1, 2)), MultinomialNB(alpha=0.1)),
        "LR so'z+bi": make_pipeline(soz((1, 2)), lr()),
        "SVC so'z+bi": make_pipeline(soz((1, 2)),
                                     LinearSVC(C=0.5, max_iter=5000)),
        "LR char_wb": make_pipeline(belgi("char_wb"), lr()),
        "LR char": make_pipeline(belgi("char"), lr()),
    }
    cv = StratifiedKFold(5, shuffle=True, random_state=0)

    natija, lugat = {}, {}
    with warnings.catch_warnings():
        # LinearSVC ba'zi foldlarda to'liq yaqinlashmasligi mumkin;
        # natijaga ta'siri yo'q, ogohlantirish chiqishni ifloslamasin
        warnings.simplefilter("ignore", ConvergenceWarning)
        for nom, m in modellar.items():
            r = cross_validate(m, matnlar, y, cv=cv, return_estimator=True)
            natija[nom] = r["test_score"]
            lugat[nom] = len(r["estimator"][0][0].vocabulary_)

    print("=== 1. 5-fold CV aniqligi (bir xil foldlar) ===")
    print(f"  {'model':<13} {'lug_at':>7} {'o_rtacha':>9} {'min':>7} "
          f"{'max':>7}")
    for nom, s in natija.items():
        print(f"  {nom:<13} {lugat[nom]:>7} {s.mean():>9.4f} "
              f"{s.min():>7.4f} {s.max():>7.4f}")
    print("  (lug'at - 1-fold o'quv qismidagi belgilar soni)")

    print("\n=== 2. Eng yaxshisi bilan juftlashgan farq ===")
    eng = max(natija, key=lambda k: natija[k].mean())
    print(f"  eng yaxshi: {eng}")
    print(f"  {'model':<13} {'farq':>8} {'SE':>7} {'xulosa':>22}")
    tanlov = None
    for nom, s in natija.items():
        if nom == eng:
            print(f"  {nom:<13} {'-':>8} {'-':>7} {'eng yaxshi':>22}")
            if tanlov is None:
                tanlov = nom
            continue
        d = s - natija[eng]
        se = d.std(ddof=1) / np.sqrt(len(d))
        yomon = d.mean() < -2 * se
        xulosa = "sezilarli yomon" if yomon else "farq sezilarli emas"
        print(f"  {nom:<13} {d.mean():>+8.4f} {se:>7.4f} {xulosa:>22}")
        if not yomon and tanlov is None:
            tanlov = nom
    print(f"\n  QAROR (sezilarli yomon bo'lmagan eng sodda): {tanlov}")

    print("\n=== 3. Bigram va belgi n-gramlari nima qo'shdi ===")
    for a, b in [("LR so'z", "LR so'z+bi"), ("NB so'z", "NB so'z+bi"),
                 ("LR char_wb", "LR char")]:
        d = natija[b] - natija[a]
        se = d.std(ddof=1) / np.sqrt(len(d))
        belgi_ = "sezilarli" if abs(d.mean()) > 2 * se else "sezilarli emas"
        print(f"  {b:<11} - {a:<11} {d.mean():>+8.4f}  SE {se:.4f}  {belgi_}")
    print("  ⭐ Qaror jadvaldan kod bilan chiqdi - oldindan yozilmagan")


if __name__ == "__main__":
    main()

=== 1. 5-fold CV aniqligi (bir xil foldlar) ===
  model          lug_at  o_rtacha     min     max
  NB so'z           861    0.6627  0.6533  0.6767
  LR so'z           861    0.7197  0.6967  0.7383
  NB so'z+bi       4619    0.7667  0.7533  0.7783
  LR so'z+bi       4619    0.8073  0.7950  0.8300
  SVC so'z+bi      4619    0.8123  0.8000  0.8283
  LR char_wb       4155    0.7327  0.7100  0.7450
  LR char          5638    0.8287  0.8133  0.8350
  (lug'at - 1-fold o'quv qismidagi belgilar soni)

=== 2. Eng yaxshisi bilan juftlashgan farq ===
  eng yaxshi: LR char
  model             farq      SE                 xulosa
  NB so'z        -0.1660  0.0056        sezilarli yomon
  LR so'z        -0.1090  0.0046        sezilarli yomon
  NB so'z+bi     -0.0620  0.0053        sezilarli yomon
  LR so'z+bi     -0.0213  0.0056        sezilarli yomon
  SVC so'z+bi    -0.0163  0.0041        sezilarli yomon
  LR char_wb     -0.0960  0.0034        sezilarli yomon
  LR char              -       -             eng yaxshi

  QAROR (sezilarli yomon bo'lmagan eng sodda): LR char

=== 3. Bigram va belgi n-gramlari nima qo'shdi ===
  LR so'z+bi  - LR so'z      +0.0877  SE 0.0064  sezilarli
  NB so'z+bi  - NB so'z      +0.1040  SE 0.0034  sezilarli
  LR char     - LR char_wb   +0.0960  SE 0.0034  sezilarli
  ⭐ Qaror jadvaldan kod bilan chiqdi - oldindan yozilmagan

"""Test sharhlariga ko'proq imlo xatosi qo'shsak: so'z va belgi modellari."""

import numpy as np
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import StratifiedKFold
from sklearn.pipeline import make_pipeline

IJOBIY = ["yaxshi", "zo'r", "a'lo", "ajoyib", "sifatli", "qulay",
          "chiroyli", "mustahkam"]
SALBIY = ["yomon", "sifatsiz", "noqulay", "xunuk", "nosoz", "mo'rt",
          "sust", "chatoq"]
NEYTRAL = ["o'rtacha", "oddiy", "normal", "odatiy", "kutilgandek"]
JIHAT = ["sifati", "narxi", "ekrani", "batareyasi", "ovozi", "qadoqlash",
         "yetkazib berish", "sotuvchi", "dizayni", "materiali"]
MAHSULOT = ["telefon", "noutbuk", "quloqchin", "soat", "choynak", "kurtka",
            "krossovka", "sumka", "planshet", "kolonka"]
BOSH = ["kecha", "o'tgan hafta", "bir oy oldin", "do'kondan", "internetdan",
        "sovg'a uchun", "akam uchun", ""]
FEL = ["oldim", "buyurtma qildim", "sotib oldim", "ishlatib ko'rdim"]
YAKUN = {1: ["tavsiya qilaman", "yana olaman", "hammaga maslahat beraman"],
         -1: ["tavsiya qilmayman", "boshqa olmayman", "pulimga achinaman"],
         0: ["hozircha shu", "vaqt ko'rsatadi", "yana ishlatib ko'ramiz"]}
ULUSH = {1: [0.8, 0.1, 0.1], -1: [0.1, 0.8, 0.1], 0: [0.15, 0.15, 0.7]}
TP = r"[\w']+"


def gap(qutb, rng):
    j = rng.choice(JIHAT)
    juda = "juda " if rng.random() < 0.3 else ""
    r = rng.random()
    if qutb == 1:
        if r < 0.2:
            return f"{j} {rng.choice(SALBIY)} emas"
        if r < 0.3:
            return f"{j} bilan muammo yo'q"
        return f"{j} {juda}{rng.choice(IJOBIY)}"
    if qutb == -1:
        if r < 0.25:
            return f"{j} {rng.choice(IJOBIY)} emas"
        if r < 0.35:
            return f"{j} bilan muammo bor"
        return f"{j} {juda}{rng.choice(SALBIY)}"
    return f"{j} {rng.choice(NEYTRAL)}"


def sharh(rng):
    kayfiyat = int(rng.integers(-1, 2))
    qismlar = [f"{rng.choice(BOSH)} {rng.choice(MAHSULOT)} "
               f"{rng.choice(FEL)}".strip()]
    ballar = []
    for t in range(int(rng.integers(1, 5))):
        q = int(rng.choice([1, -1, 0], p=ULUSH[kayfiyat]))
        bog = rng.choice(["lekin", "va", "ammo", "ham", ""]) if t else ""
        qismlar.append(f"{bog} {gap(q, rng)}".strip())
        ballar.append(q)
    if rng.random() < 0.5:
        q = int(rng.choice([1, -1, 0], p=ULUSH[kayfiyat]))
        qismlar.append(rng.choice(YAKUN[q]))
        ballar.append(q)
    ball = np.mean(ballar) + rng.normal(0, 0.1)
    return " ".join(qismlar), 2 if ball > 0.35 else (0 if ball < -0.35 else 1)


def imlo_xato(matn, ulush, rng):
    chiqish = []
    for s in matn.split():
        if len(s) >= 4 and rng.random() < ulush:
            i = int(rng.integers(1, len(s) - 1))
            tur = int(rng.integers(0, 3))
            if tur == 0:
                s = s[:i] + s[i + 1:]
            elif tur == 1:
                s = s[:i] + s[i] + s[i:]
            else:
                s = s[:i - 1] + s[i] + s[i - 1] + s[i + 1:]
        chiqish.append(s)
    return " ".join(chiqish)


def korpus(n, seed=0, xato=0.07):
    rng = np.random.default_rng(seed)
    toza, y = zip(*[sharh(rng) for _ in range(n)])
    rng2 = np.random.default_rng(seed + 1000)
    return [imlo_xato(t, xato, rng2) for t in toza], np.array(y), list(toza)


def modellar():
    lr = lambda: LogisticRegression(C=10, max_iter=3000)
    soz = lambda: TfidfVectorizer(token_pattern=TP, ngram_range=(1, 2))
    belgi = lambda: TfidfVectorizer(analyzer="char", ngram_range=(2, 4))
    return {"so'z+bigram": make_pipeline(soz(), lr()),
            "belgi (char)": make_pipeline(belgi(), lr())}


def main() -> None:
    matnlar, y, toza = korpus(3000)
    darajalar = [0.0, 0.15, 0.3]
    cv = StratifiedKFold(5, shuffle=True, random_state=0)
    nomlar = list(modellar())
    natija = {n: {d: [] for d in darajalar} for n in nomlar}
    oov = {d: [] for d in darajalar}

    for k, (tr, te) in enumerate(cv.split(matnlar, y)):
        oquv = [matnlar[i] for i in tr]
        lugat = {s for m in oquv for s in m.split()}
        testlar = {}
        for d in darajalar:
            rng = np.random.default_rng(100 + k)
            testlar[d] = [imlo_xato(toza[i], d, rng) for i in te]
            sozlar = [s for m in testlar[d] for s in m.split()]
            oov[d].append(np.mean([s not in lugat for s in sozlar]))
        for nom, m in modellar().items():
            m.fit(oquv, y[tr])
            for d in darajalar:
                natija[nom][d].append(np.mean(m.predict(testlar[d]) == y[te]))

    print("=== 1. Test so'zlarining lug'atdan tashqari (OOV) ulushi ===")
    for d in darajalar:
        print(f"  xato darajasi {d:.2f}: OOV {np.mean(oov[d]):.1%}")

    print("\n=== 2. Aniqlik (5 fold o'rtachasi) ===")
    print(f"  {'xato':>5} " + " ".join(f"{n:>13}" for n in nomlar))
    for d in darajalar:
        print(f"  {d:>5.2f} " + " ".join(f"{np.mean(natija[n][d]):>13.4f}"
                                         for n in nomlar))

    print("\n=== 3. Aniqlik yo'qotishi (0.0 dan 0.3 gacha) ===")
    for n in nomlar:
        a0 = np.mean(natija[n][0.0])
        a4 = np.mean(natija[n][0.3])
        print(f"  {n:<13} {a0:.4f} -> {a4:.4f}  "
              f"(yo'qotish {a0 - a4:.4f})")

    print("\n=== 4. Juftlashgan farq: belgi - so'z+bigram ===")
    print(f"  {'xato':>5} {'farq':>8} {'SE':>7} {'xulosa':>22}")
    for d in darajalar:
        f = (np.array(natija["belgi (char)"][d])
             - np.array(natija["so'z+bigram"][d]))
        se = f.std(ddof=1) / np.sqrt(len(f))
        if abs(f.mean()) <= 2 * se:
            xulosa = "sezilarli emas"
        elif f.mean() > 0:
            xulosa = "belgi sezilarli yaxshi"
        else:
            xulosa = "so'z sezilarli yaxshi"
        print(f"  {d:>5.2f} {f.mean():>+8.4f} {se:>7.4f} {xulosa:>22}")
    y0 = np.mean(natija["so'z+bigram"][0.0]) - np.mean(
        natija["so'z+bigram"][0.3])
    y1 = np.mean(natija["belgi (char)"][0.0]) - np.mean(
        natija["belgi (char)"][0.3])
    if y1 < y0:
        print(f"  belgi modeli {y0 / y1:.1f} barobar kam yo'qotdi")
    else:
        print("  belgi modeli kamroq yo'qotmadi")
    print("  ⭐ Imlo xatosi ko'p bo'lsa - belgi n-gramlari")


if __name__ == "__main__":
    main()

=== 1. Test so'zlarining lug'atdan tashqari (OOV) ulushi ===
  xato darajasi 0.00: OOV 0.0%
  xato darajasi 0.15: OOV 1.9%
  xato darajasi 0.30: OOV 3.8%

=== 2. Aniqlik (5 fold o'rtachasi) ===
   xato   so'z+bigram  belgi (char)
   0.00        0.8320        0.8367
   0.15        0.7883        0.8113
   0.30        0.7267        0.7863

=== 3. Aniqlik yo'qotishi (0.0 dan 0.3 gacha) ===
  so'z+bigram   0.8320 -> 0.7267  (yo'qotish 0.1053)
  belgi (char)  0.8367 -> 0.7863  (yo'qotish 0.0503)

=== 4. Juftlashgan farq: belgi - so'z+bigram ===
   xato     farq      SE                 xulosa
   0.00  +0.0047  0.0051         sezilarli emas
   0.15  +0.0230  0.0029 belgi sezilarli yaxshi
   0.30  +0.0597  0.0083 belgi sezilarli yaxshi
  belgi modeli 2.1 barobar kam yo'qotdi
  ⭐ Imlo xatosi ko'p bo'lsa - belgi n-gramlari

"""LogReg koeffitsiyentlari, chalkashlik matritsasi va eng ishonchli xatolar."""

import numpy as np
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import confusion_matrix
from sklearn.model_selection import train_test_split
from sklearn.pipeline import make_pipeline

IJOBIY = ["yaxshi", "zo'r", "a'lo", "ajoyib", "sifatli", "qulay",
          "chiroyli", "mustahkam"]
SALBIY = ["yomon", "sifatsiz", "noqulay", "xunuk", "nosoz", "mo'rt",
          "sust", "chatoq"]
NEYTRAL = ["o'rtacha", "oddiy", "normal", "odatiy", "kutilgandek"]
JIHAT = ["sifati", "narxi", "ekrani", "batareyasi", "ovozi", "qadoqlash",
         "yetkazib berish", "sotuvchi", "dizayni", "materiali"]
MAHSULOT = ["telefon", "noutbuk", "quloqchin", "soat", "choynak", "kurtka",
            "krossovka", "sumka", "planshet", "kolonka"]
BOSH = ["kecha", "o'tgan hafta", "bir oy oldin", "do'kondan", "internetdan",
        "sovg'a uchun", "akam uchun", ""]
FEL = ["oldim", "buyurtma qildim", "sotib oldim", "ishlatib ko'rdim"]
YAKUN = {1: ["tavsiya qilaman", "yana olaman", "hammaga maslahat beraman"],
         -1: ["tavsiya qilmayman", "boshqa olmayman", "pulimga achinaman"],
         0: ["hozircha shu", "vaqt ko'rsatadi", "yana ishlatib ko'ramiz"]}
ULUSH = {1: [0.8, 0.1, 0.1], -1: [0.1, 0.8, 0.1], 0: [0.15, 0.15, 0.7]}
SINFLAR = ["salbiy", "neytral", "ijobiy"]
TP = r"[\w']+"


def gap(qutb, rng):
    j = rng.choice(JIHAT)
    juda = "juda " if rng.random() < 0.3 else ""
    r = rng.random()
    if qutb == 1:
        if r < 0.2:
            return f"{j} {rng.choice(SALBIY)} emas"
        if r < 0.3:
            return f"{j} bilan muammo yo'q"
        return f"{j} {juda}{rng.choice(IJOBIY)}"
    if qutb == -1:
        if r < 0.25:
            return f"{j} {rng.choice(IJOBIY)} emas"
        if r < 0.35:
            return f"{j} bilan muammo bor"
        return f"{j} {juda}{rng.choice(SALBIY)}"
    return f"{j} {rng.choice(NEYTRAL)}"


def sharh(rng):
    kayfiyat = int(rng.integers(-1, 2))
    qismlar = [f"{rng.choice(BOSH)} {rng.choice(MAHSULOT)} "
               f"{rng.choice(FEL)}".strip()]
    ballar = []
    for t in range(int(rng.integers(1, 5))):
        q = int(rng.choice([1, -1, 0], p=ULUSH[kayfiyat]))
        bog = rng.choice(["lekin", "va", "ammo", "ham", ""]) if t else ""
        qismlar.append(f"{bog} {gap(q, rng)}".strip())
        ballar.append(q)
    if rng.random() < 0.5:
        q = int(rng.choice([1, -1, 0], p=ULUSH[kayfiyat]))
        qismlar.append(rng.choice(YAKUN[q]))
        ballar.append(q)
    ball = np.mean(ballar) + rng.normal(0, 0.1)
    return " ".join(qismlar), 2 if ball > 0.35 else (0 if ball < -0.35 else 1)


def imlo_xato(matn, ulush, rng):
    chiqish = []
    for s in matn.split():
        if len(s) >= 4 and rng.random() < ulush:
            i = int(rng.integers(1, len(s) - 1))
            tur = int(rng.integers(0, 3))
            if tur == 0:
                s = s[:i] + s[i + 1:]
            elif tur == 1:
                s = s[:i] + s[i] + s[i:]
            else:
                s = s[:i - 1] + s[i] + s[i - 1] + s[i + 1:]
        chiqish.append(s)
    return " ".join(chiqish)


def korpus(n, seed=0, xato=0.07):
    rng = np.random.default_rng(seed)
    toza, y = zip(*[sharh(rng) for _ in range(n)])
    rng2 = np.random.default_rng(seed + 1000)
    return [imlo_xato(t, xato, rng2) for t in toza], np.array(y), list(toza)


def main() -> None:
    matnlar, y, _ = korpus(3000)
    Xtr, Xte, ytr, yte = train_test_split(matnlar, y, test_size=0.2,
                                          random_state=0, stratify=y)
    model = make_pipeline(
        TfidfVectorizer(token_pattern=TP, ngram_range=(1, 2)),
        LogisticRegression(C=10, max_iter=3000))
    model.fit(Xtr, ytr)
    vek, clf = model[0], model[1]
    nomlar = vek.get_feature_names_out()

    print("=== 1. Har sinf uchun eng kuchli belgilar ===")
    for k in (2, 0, 1):
        top = np.argsort(clf.coef_[k])[::-1][:7]
        print(f"  {SINFLAR[k]}:")
        print("    " + ", ".join(f"{nomlar[i]} ({clf.coef_[k, i]:.1f})"
                                 for i in top[:4]))
        print("    " + ", ".join(f"{nomlar[i]} ({clf.coef_[k, i]:.1f})"
                                 for i in top[4:]))

    print("\n=== 2. Inkor: bir so'z, ikki ma'no ===")
    j = {n: i for i, n in enumerate(nomlar)}
    for f in ["sifatli", "sifatli emas", "sust", "sust emas", "emas"]:
        if f in j:
            w = clf.coef_[2, j[f]] - clf.coef_[0, j[f]]
            print(f"  {f:<14} ijobiy - salbiy og'irlik: {w:+.2f}")
    print("  unigram 'emas' deyarli neytral - ma'noni bigram tashiydi")

    print("\n=== 3. Shubhali belgilar: mahsulot nomlari ===")
    kuch = np.abs(clf.coef_).max(0)
    tartib = np.argsort(-kuch)
    rang = {nomlar[i]: r for r, i in enumerate(tartib)}
    for m in ["telefon", "kurtka", "soat"]:
        print(f"  {m:<9} eng katta |og'irlik| {kuch[j[m]]:.2f}, "
              f"o'rni {rang[m] + 1} / {len(nomlar)}")
    print("  mahsulot nomi sinfga bog'liq emas - og'irligi kichik bo'lishi kerak")

    print("\n=== 4. Chalkashlik matritsasi (qator - haqiqiy) ===")
    p = model.predict_proba(Xte)
    b = p.argmax(1)
    M = confusion_matrix(yte, b)
    print("           " + "".join(f"{s:>9}" for s in SINFLAR))
    for i, s in enumerate(SINFLAR):
        print(f"  {s:<8} " + "".join(f"{M[i, k]:>9}" for k in range(3)))
    print(f"  aniqlik: {np.trace(M) / M.sum():.4f}")
    print(f"  salbiy <-> ijobiy xatolar: {M[0, 2] + M[2, 0]}, "
          f"neytral bilan bog'liq: {M.sum() - np.trace(M) - M[0, 2] - M[2, 0]}")

    print("\n=== 5. Guruhlar bo'yicha aniqlik ===")
    togri = b == yte
    inkor = np.array([" emas" in m for m in Xte])
    gaplar = np.array([sum(m.count(f" {c} ") + m.startswith(c)
                           for c in JIHAT) for m in Xte])
    for nom, g in [("inkorli", inkor), ("inkorsiz", ~inkor),
                   ("1 jihat", gaplar <= 1), ("3+ jihat", gaplar >= 3)]:
        print(f"  {nom:<9} {int(g.sum()):>4} ta, aniqlik {togri[g].mean():.4f}")

    print("\n=== 6. Eng ishonchli xatolar ===")
    xato = np.where(~togri)[0]
    ishonch = p[xato, b[xato]]
    for i in xato[np.argsort(-ishonch)[:3]]:
        print(f"  haqiqiy {SINFLAR[yte[i]]}, bashorat {SINFLAR[b[i]]} "
              f"(p={p[i, b[i]]:.2f})")
        print(f"    {Xte[i]}")
    print(f"  xatolardagi o'rtacha ishonch: {ishonch.mean():.3f}, "
          f"to'g'rilarda: {p[togri, b[togri]].mean():.3f}")
    print("  ⭐ Koeffitsiyent va xatolar - modelni tushunish yo'li")


if __name__ == "__main__":
    main()

=== 1. Har sinf uchun eng kuchli belgilar ===
  ijobiy:
    nosoz emas 7.3-bob, noqulay emas 6.4-bob, yomon emas 5.7-bob, mo'rt emas 5.6-bob
    sust emas 5.5-bob, sifatsiz emas 5.1-bob, zo'r 4.9-bob
  salbiy:
    mustahkam emas 6.3-bob, zo'r emas 5.9-bob, sifatli emas 5.7-bob, yaxshi emas 5.1-bob
    chatoq 5.0-bob, chiroyli emas 4.9-bob, bor 4.9-bob
  neytral:
    odatiy 5.7-bob, o'rtacha 5.3-bob, oddiy 5.1-bob, kutilgandek 5.0-bob
    normal 4.4-bob, ekrani odatiy 3.3-bob, lekin qadoqlash 2.8-bob

=== 2. Inkor: bir so'z, ikki ma'no ===
  sifatli        ijobiy - salbiy og'irlik: +7.12
  sifatli emas   ijobiy - salbiy og'irlik: -11.01
  sust           ijobiy - salbiy og'irlik: -6.77
  sust emas      ijobiy - salbiy og'irlik: +10.50
  emas           ijobiy - salbiy og'irlik: -2.25
  unigram 'emas' deyarli neytral - ma'noni bigram tashiydi

=== 3. Shubhali belgilar: mahsulot nomlari ===
  telefon   eng katta |og'irlik| 0.63, o'rni 1470 / 4612
  kurtka    eng katta |og'irlik| 0.44, o'rni 2177 / 4612
  soat      eng katta |og'irlik| 0.83, o'rni 953 / 4612
  mahsulot nomi sinfga bog'liq emas - og'irligi kichik bo'lishi kerak

=== 4. Chalkashlik matritsasi (qator - haqiqiy) ===
              salbiy  neytral   ijobiy
  salbiy         153       28        7
  neytral         28      156       24
  ijobiy           2       22      180
  aniqlik: 0.8150
  salbiy <-> ijobiy xatolar: 9, neytral bilan bog'liq: 102

=== 5. Guruhlar bo'yicha aniqlik ===
  inkorli    190 ta, aniqlik 0.7684
  inkorsiz   410 ta, aniqlik 0.8366
  1 jihat    171 ta, aniqlik 0.8538
  3+ jihat   283 ta, aniqlik 0.8269

=== 6. Eng ishonchli xatolar ===
  haqiqiy neytral, bashorat ijobiy (p=0.99)
    noutbuk sotib oldim materiali chiroyli ammo sifati mo'rt qadoqlash juda a'lo yana olaman
  haqiqiy ijobiy, bashorat neytral (p=0.98)
    bir oy oldin sumka oldim batareyasi kutilgandek lekin sifati chatoq emas ammo ovozi kutilgandek ammo sifati kutilgandek
  haqiqiy neytral, bashorat salbiy (p=0.97)
    akam uchun krossovka buyurtma qildim narxi juda mo'rt ham batareeyasi kutilgandek ham ekrani zo'r emas lekin sifati juuda sifatsiz
  xatolardagi o'rtacha ishonch: 0.681, to'g'rilarda: 0.847
  ⭐ Koeffitsiyent va xatolar - modelni tushunish yo'li

X = TfidfVectorizer().fit_transform(matnlar); train_test_split(X, y)  # ⚠️
model = make_pipeline(TfidfVectorizer(), LogisticRegression())       # ✅

TfidfVectorizer()                            # "zo'r" -> "zo"      # ⚠️
TfidfVectorizer(token_pattern=r"[\w']+")     # "zo'r" butun        # ✅

TfidfVectorizer(ngram_range=(1, 1))          # "emas" alohida      # ⚠️
TfidfVectorizer(ngram_range=(1, 2))          # "sifatli emas"      # ✅

cross_val_score(a, X, y, cv=5); cross_val_score(b, X, y, cv=KFold(5, shuffle=True))  # ⚠️
cv = StratifiedKFold(5, shuffle=True, random_state=0)  # ikkalasiga bir xil  # ✅

make_pipeline(TfidfVectorizer(), StandardScaler(with_mean=False), MultinomialNB())  # ⚠️ o'lchov buziladi
make_pipeline(TfidfVectorizer(), MultinomialNB())  # sanoq/TF-IDF - manfiy emas     # ✅

LinearSVC().predict_proba(X)                 # AttributeError      # ⚠️
LogisticRegression().predict_proba(X)        # yoki CalibratedClassifierCV  # ✅

print("char yaxshi" if aniq_char > aniq_soz else "so'z yaxshi")    # ⚠️
d = char_s - soz_s; abs(d.mean()) > 2 * d.std(ddof=1) / len(d) ** 0.5  # ✅

1.  # 3 ta teng sinfda DummyClassifier aniqligi taxminan?
2.  # TfidfVectorizer ning fit qismi nimalarni o'rganadi?
3.  # sukut token_pattern "zo'r" ni qanday bo'ladi?
4.  # "sifatli emas" ni unigram modeli qanday ko'radi?
5.  # analyzer="char_wb" va "char" farqi?
6.  # MultinomialNB dagi alpha nima uchun?
7.  # LinearSVC predict_proba beradimi?
8.  # juftlashgan CV da nima bir xil bo'lishi shart?
9.  # SE formulasi (5 fold)?
10. # qaror qoidasi: qaysi model tanlanadi?
11. # clf.coef_ shakli (3 sinf, 20000 belgi)?
12. # mahsulot nomi katta og'irlik olsa - nima belgisi?

1.  X = TfidfVectorizer().fit_transform(matnlar)
    Xtr, Xte, ytr, yte = train_test_split(X, y)

2.  vek = TfidfVectorizer()          # o'zbekcha sharhlar uchun

3.  make_pipeline(TfidfVectorizer(), StandardScaler(with_mean=False),
                  MultinomialNB())

4.  a = cross_val_score(m1, X, y, cv=KFold(5, shuffle=True))
    b = cross_val_score(m2, X, y, cv=KFold(5, shuffle=True))

5.  p = make_pipeline(TfidfVectorizer(), LinearSVC()).fit(X, y).predict_proba(X2)