Python kursi25.5-dars: Feature engineering
Python kursi · ML va AI

25.5-dars: Feature engineering

16 daqiqa5/12-dars

25-QISM — ML VA AI · 5-dars


1. Kirish va motivatsiya

25.3 da ma'lumotni tayyorladik (kodlash, masshtablash). Endi kuchliroq mahorat: mavjud ma'lumotdan yangi, foydali xususiyatlar yaratish. Masalan: sana o'zi kam foyda, lekin undan "hafta kuni", "oymi/dam olish kunimi" chiqarish model uchun qimmatli. Narx va soni alohida emas, "jami summa" (narx × soni) muhimroq. Bu — feature engineering (xususiyat muhandisligi).

Feature engineering — mavjud ma'lumotdan model uchun foydaliroq xususiyatlar yaratish: hosila (narx × soni = jami), guruhlash (yosh → yosh guruhi), sana qismlari (24.11 — hafta kuni, oy), matn (so'zlar → sonlar), polinom (x → x^2, x×y). Ko'pincha feature engineering model tanlovidan ko'proq ta'sir qiladi — "yaxshi feature + oddiy model > yomon feature + murakkab model". Bu ML ning ijodiy va muhim qismi. To'g'ri feature — model muvaffaqiyati.

Real vaziyat. Bir model uy narxini bashorat qilardi — o'rtacha aniqlik. Sana ustuni bor edi (qurilgan yili). Undan "yoshi" (2024 - yil) feature yaratildi — bu ma'noliroq (eski uy arzon). Va "narx/maydon" (kvadrat metr narxi) qo'shildi. Aniqlik sezilarli oshdi. Yangi feature model tanlovidan ko'proq yordam berdi. Feature engineering — ML ning yashirin kuchi.

Bu darsda feature engineering'ni o'rganamiz.

Bu darsda:

  • Feature engineering nima
  • Hosila feature (kombinatsiya)
  • Guruhlash (binning)
  • Sana feature (24.11)
  • Matn feature (so'zlar → son)
  • Polinom feature
  • Feature tanlovi
  • Amaliy: feature yaratish

ℹ Misollarda scikit-learn, pandas bilan sinaladi.


2. Nazariya — chuqur tushuntirish

2.1. Feature engineering nima

Ma'lumotdan foydaliroq xususiyatlar:

Xom: sana="2020-01-15", narx=100, soni=5
Yangi feature:
   yosh = 2024 - 2020 = 4        (sanadan)
   jami = narx * soni = 500      (kombinatsiya)
   hafta_kuni = ...              (sanadan)

Feature engineering — mavjud ma'lumotdan model uchun foydaliroq xususiyatlar yaratish: hosila (narx × soni), sana qismlari (hafta kuni), guruhlash (yosh → guruh), nisbat (narx / maydon). Maqsad: modelga ma'lumotdagi naqshni oson ko'rsatish (model o'zi topmasa ham). "Yaxshi feature — yarim model" — ko'pincha model tanlovidan muhimroq. Bilim va ijod talab qiladi (soha bilimi).

2.2. Hosila feature (kombinatsiya)

Mavjud ustunlardan yangi:

python
df["jami"] = df["narx"] * df["soni"]          # ko'paytma
df["narx_metr"] = df["narx"] / df["maydon"]   # nisbat
df["farq"] = df["daromad"] - df["xarajat"]    # ayirma
df["narx_log"] = np.log1p(df["narx"])         # logarifm

Hosila feature — mavjud ustunlardan matematik yangi: ko'paytma (narx × soni = jami), nisbat (narx / maydon = metr narxi — ko'pincha muhim), ayirma (daromad - xarajat = foyda), logarifm (log(narx) — katta oraliqni siqadi, egri taqsimotni normal). Bu ma'lumotdagi ma'noli bog'lanishni ochadi. Soha bilimi kerak (nima muhim — narx yoki metr narxi?). Eng ko'p ishlatiladigan feature turi.

2.3. Guruhlash (binning)

Uzluksiz sonni toifaga:

python
df["yosh_guruh"] = pd.cut(
    df["yosh"], bins=[0, 18, 35, 60, 100],
    labels=["bola", "yosh", "o'rta", "keksa"]
)
# 25 → "yosh", 45 → "o'rta"

Guruhlash (binning) — uzluksiz sonni toifaga bo'lish (pd.cut — 24.11): yosh (son) → yosh guruhi (bola/yosh/o'rta/keksa). Sabab: ba'zan toifa son'dan ma'noliroq (25 va 30 yosh o'xshash — "yosh" guruh), chiziqsiz bog'lanish (yosh va xarid — chiziqli emas). pd.qcut (teng sonli guruhlar). Guruhlash ma'lumotni soddalashtiradi (ba'zan yaxshi, ba'zan ma'lumot yo'qotadi).

2.4. Sana feature (24.11)

Sanadan foydali qismlar:

python
df["oy"] = df["sana"].dt.month           # oy (mavsumiylik)
df["hafta_kuni"] = df["sana"].dt.dayofweek  # 0-6
df["dam_olish"] = df["hafta_kuni"] >= 5   # dam olish kuni
df["yosh"] = 2024 - df["yil"]            # yosh

Sana feature 24.11-bob — sana o'zi kam foyda, qismlari qimmatli: dt.month (mavsumiylik — dekabrda xarid ko'p), dt.dayofweek (hafta kuni — dam olish kuni sotuv boshqacha), dt.hour (soat), dam_olish (bool). Yoki "yosh" (hozir - sana). Model sanani tushunmaydi (raqam), lekin "dekabrmi?", "dam olishmi?" — ma'noli feature. Vaqt ma'lumotdan foydali feature.

2.5. Matn feature (so'zlar → son)

Matnni songa (model son kutadi):

python
from sklearn.feature_extraction.text import CountVectorizer
cv = CountVectorizer()
X = cv.fit_transform(["python kod", "kod yaxshi"])
# har so'z → ustun, qiymat = chastota
# TfidfVectorizer — muhimlik bilan

Matn feature — matnni songa (model matn tushunmaydi): CountVectorizer (har so'z → ustun, qiymat = so'z chastotasi — "bag of words"), TfidfVectorizer (chastota × noyoblik — muhim so'zlarga ko'proq vazn). Matn (sharh, xat) → son matritsa → model. Bu matn klassifikatsiyasi (spam, hissiyot) asosi. Zamonaviy usul — embedding (25.10 LLM). Matnni ML uchun tayyorlash.

2.6. Polinom feature

Chiziqsiz bog'lanish uchun:

python
from sklearn.preprocessing import PolynomialFeatures
poly = PolynomialFeatures(degree=2, include_bias=False)
poly.fit_transform(X)
# [a, b] → [a, b, a^2, a*b, b^2]

Polinom feature (PolynomialFeatures) — mavjud featuredan daraja va ko'paytma: [a, b] → [a, b, a^2, a×b, b^2]. Sabab: chiziqli model (regressiya) faqat chiziqli bog'lanish topadi — polinom feature chiziqsizni (egri) o'rgatadi. a×b — o'zaro ta'sir (interaction — maydon va joylashuv birga). Ehtiyot: ko'p feature (overfitting, 25.2). Chiziqsiz naqshni chiziqli modelga.

2.7. Feature tanlovi

Ko'p feature — har doim yaxshi emas:

python
from sklearn.feature_selection import SelectKBest
# eng muhim feature'larni tanla
# yoki model.feature_importances_ (25.2)

Feature tanlovi — eng muhim featurelarni tanlash (hammani emas): ko'p feature — overfitting 25.2-bob, sekin, shovqin (foydasiz feature). SelectKBest (eng yaxshi K), feature_importances_ (model — qaysi muhim). "Kam, lekin yaxshi feature" — ko'p, shovqinli'dan afzal. Bu "Occam ustarasi" (soddalik) — keraksiz feature model'ni buzadi. Feature yaratish + tanlov — muvozanat.

2.8. Feature engineering — ijod va bilim

Feature engineering — ML ning ijodiy va ko'pincha eng ta'sirli qismi: yaxshi feature + oddiy model > yomon feature + murakkab model. Soha bilimi kerak (uy narxida metr narxi muhim — buni bilish kerak). Model ma'lumotdagi naqshni topadi, lekin feature naqshni oson ko'rsatadi (model o'zi topmasa ham). Zamonaviy chuqur o'qitish 25.8-bob feature'ni avtomatik o'rganadi, lekin klassik ML'da feature engineering hal qiladi. "Ma'lumotni tushunish — model'dan muhimroq".


3. Tez ma'lumotnoma

python
import pandas as pd, numpy as np
from sklearn.preprocessing import PolynomialFeatures
from sklearn.feature_extraction.text import CountVectorizer

# hosila feature:
df["jami"] = df["narx"] * df["soni"]
df["nisbat"] = df["narx"] / df["maydon"]
df["log_narx"] = np.log1p(df["narx"])

# guruhlash (binning):
df["guruh"] = pd.cut(df["yosh"], bins=[0,18,35,60,100], labels=[...])

# sana feature 24.11-bob:
df["oy"] = df["sana"].dt.month
df["hafta_kuni"] = df["sana"].dt.dayofweek
df["dam_olish"] = df["hafta_kuni"] >= 5

# matn feature:
CountVectorizer().fit_transform(matnlar)

# polinom:
PolynomialFeatures(degree=2, include_bias=False).fit_transform(X)

Feature engineering xulosasi

Yangi, foydali xususiyatlar · hosila (narx*soni) / nisbat (narx/maydon)
Binning (son→toifa) / sana (oy, hafta_kuni) / matn (so'z→son) / polinom
Yaxshi feature + oddiy model > yomon feature + murakkab · soha bilimi

4. Batafsil misollar

Misollarda scikit-learn, pandas bilan sinaladi.

Misol 1 — Hosila feature

python
"""hosila feature: ko'paytma (jami), nisbat (metr narxi), logarifm — ma'noli kombinatsiya."""

import warnings
warnings.filterwarnings("ignore")

import numpy as np
import pandas as pd


def main() -> None:
    df = pd.DataFrame({
        "narx": [100, 200, 300, 150],
        "soni": [5, 3, 2, 8],
        "maydon": [50, 80, 100, 60],
    })

    print("=== 1. Ko'paytma (jami summa) ===")
    df["jami"] = df["narx"] * df["soni"]
    print(f"  jami: {df['jami'].tolist()}")

    print("\n=== 2. Nisbat (metr narxi) ===")
    df["narx_metr"] = (df["narx"] / df["maydon"]).round(2)
    print(f"  narx/maydon: {df['narx_metr'].tolist()}")

    print("\n=== 3. Logarifm (katta oraliqni siqash) ===")
    df["log_jami"] = np.log1p(df["jami"]).round(3)
    print(f"  log(jami): {df['log_jami'].tolist()}")

    print("\n=== 4. Nega hosila feature ===")
    print("  jami — narx va sonidan ma'noliroq")
    print("  metr narxi — uylarni solishtirishda adolatli")
    print("  ⭐ hosila feature — ma'lumotdagi ma'noli bog'lanish")


if __name__ == "__main__":
    main()

Natijaning muhim qismi:

text
=== 1. Ko'paytma (jami summa) ===
  jami: [500, 600, 600, 1200]

=== 2. Nisbat (metr narxi) ===
  narx/maydon: [2.0, 2.5, 3.0, 2.5]

=== 3. Logarifm (katta oraliqni siqash) ===
  log(jami): [6.217, 6.399, 6.399, 7.091]

=== 4. Nega hosila feature ===
  jami — narx va sonidan ma'noliroq
  metr narxi — uylarni solishtirishda adolatli
  ⭐ hosila feature — ma'lumotdagi ma'noli bog'lanish

Nima ko'rsatdi: 2.2-bo'lim.

Misol 2 — Guruhlash va sana feature

python
"""binning (pd.cut — son → toifa); sana feature (dt.month, dt.dayofweek, dam_olish)."""

import warnings
warnings.filterwarnings("ignore")

import pandas as pd


def main() -> None:
    print("=== 1. Guruhlash (binning) ===")
    df = pd.DataFrame({"yosh": [15, 25, 35, 45, 65]})
    df["guruh"] = pd.cut(
        df["yosh"], bins=[0, 18, 35, 60, 100],
        labels=["bola", "yosh", "o'rta", "keksa"],
    )
    print(f"  yosh: {df['yosh'].tolist()}")
    print(f"  guruh: {df['guruh'].tolist()}")

    print("\n=== 2. Sana feature ===")
    df2 = pd.DataFrame({"sana": pd.to_datetime(["2024-01-15", "2024-06-20", "2024-12-25"])})
    df2["oy"] = df2["sana"].dt.month
    df2["hafta_kuni"] = df2["sana"].dt.dayofweek
    print(f"  oylar: {df2['oy'].tolist()}")
    print(f"  hafta kuni: {df2['hafta_kuni'].tolist()}")

    print("\n=== 3. Dam olish kuni (bool feature) ===")
    df2["dam_olish"] = df2["hafta_kuni"] >= 5
    print(f"  dam olish: {df2['dam_olish'].tolist()}")

    print("\n=== 4. Nega bu feature'lar ===")
    print("  yosh guruh — son'dan ma'noliroq (25≈30)")
    print("  oy — mavsumiylik, dam olish — xarid naqshi")
    print("  ⭐ binning (son→toifa), sana (oy/hafta) — ma'noli feature")


if __name__ == "__main__":
    main()

Natijaning muhim qismi:

text
=== 1. Guruhlash (binning) ===
  yosh: [15, 25, 35, 45, 65]
  guruh: ['bola', 'yosh', 'yosh', "o'rta", 'keksa']

=== 2. Sana feature ===
  oylar: [1, 6, 12]
  hafta kuni: [0, 3, 2]

=== 3. Dam olish kuni (bool feature) ===
  dam olish: [False, False, False]

=== 4. Nega bu feature'lar ===
  yosh guruh — son'dan ma'noliroq (25≈30)
  oy — mavsumiylik, dam olish — xarid naqshi
  ⭐ binning (son→toifa), sana (oy/hafta) — ma'noli feature

Nima ko'rsatdi: 2.3, 2.4-bo'limlar.

Misol 3 — Matn va polinom feature

python
"""matn feature (CountVectorizer — so'z → son); polinom feature (PolynomialFeatures)."""

import warnings
warnings.filterwarnings("ignore")

import numpy as np
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.preprocessing import PolynomialFeatures


def main() -> None:
    print("=== 1. Matn feature (CountVectorizer) ===")
    matnlar = ["python kod yaxshi", "kod python", "yaxshi dastur"]
    cv = CountVectorizer()
    X = cv.fit_transform(matnlar)
    print(f"  lug'at: {sorted(cv.vocabulary_.keys())}")
    print(f"  matritsa:")
    for m, row in zip(matnlar, X.toarray()):
        print(f"    '{m}' → {row.tolist()}")

    print("\n=== 2. Polinom feature ===")
    X2 = np.array([[2, 3], [4, 5]])
    poly = PolynomialFeatures(degree=2, include_bias=False)
    Xp = poly.fit_transform(X2)
    print(f"  nomlar: {list(poly.get_feature_names_out(['a', 'b']))}")
    print(f"  [2,3] → {Xp[0].tolist()}")

    print("\n=== 3. Nega matn feature ===")
    print("  model matn tushunmaydi → so'z chastotasi (son)")

    print("\n=== 4. Nega polinom ===")
    print("  chiziqli model + polinom → chiziqsiz naqsh (a*b, a^2)")
    print("  ⭐ matn (so'z→son), polinom (chiziqsiz naqsh)")


if __name__ == "__main__":
    main()

Natijaning muhim qismi:

text
=== 1. Matn feature (CountVectorizer) ===
  lug'at: ['dastur', 'kod', 'python', 'yaxshi']
  matritsa:
    'python kod yaxshi' → [0, 1, 1, 1]
    'kod python' → [0, 1, 1, 0]
    'yaxshi dastur' → [1, 0, 0, 1]

=== 2. Polinom feature ===
  nomlar: ['a', 'b', 'a^2', 'a b', 'b^2']
  [2,3] → [2.0, 3.0, 4.0, 6.0, 9.0]

=== 3. Nega matn feature ===
  model matn tushunmaydi → so'z chastotasi (son)

=== 4. Nega polinom ===
  chiziqli model + polinom → chiziqsiz naqsh (a*b, a^2)
  ⭐ matn (so'z→son), polinom (chiziqsiz naqsh)

Nima ko'rsatdi: 2.5, 2.6-bo'limlar.

Misol 4 — Amaliy: feature engineering va model

Real ML: xom ma'lumotga feature engineering qo'llab, model aniqligini oshirish. Bu — feature yaratishning ta'sirini ko'rsatadigan namuna.

python
"""to'liq: xom feature vs boyitilgan feature — model aniqligi (feature engineering ta'siri)."""

import warnings
warnings.filterwarnings("ignore")

import numpy as np
import pandas as pd
from sklearn.linear_model import LinearRegression


def main() -> None:
    # uy ma'lumoti
    df = pd.DataFrame({
        "maydon": [50, 70, 100, 60, 90, 80],
        "yil": [2020, 2015, 2022, 2010, 2018, 2021],
        "narx": [100, 130, 220, 110, 180, 175],
    })

    print("=== 1. Xom feature (maydon, yil) ===")
    X_xom = df[["maydon", "yil"]]
    m1 = LinearRegression().fit(X_xom, df["narx"])
    print(f"  R2 (xom): {round(m1.score(X_xom, df['narx']), 3)}")

    print("\n=== 2. Feature engineering ===")
    df["yosh"] = 2024 - df["yil"]              # yosh (yildan)
    df["narx_kutish"] = df["maydon"] * 2       # taxminiy narx
    print(f"  yangi: yosh, narx_kutish")

    print("\n=== 3. Boyitilgan feature bilan ===")
    X_boy = df[["maydon", "yosh"]]
    m2 = LinearRegression().fit(X_boy, df["narx"])
    print(f"  R2 (yosh bilan): {round(m2.score(X_boy, df['narx']), 3)}")

    print("\n=== 4. Xulosa ===")
    print(f"  'yil' o'rniga 'yosh' — ma'noliroq feature")
    print(f"  feature engineering — model tanlovidan muhim")
    print("  ⭐ yaxshi feature — model aniqligini oshiradi")


if __name__ == "__main__":
    main()

Natijaning muhim qismi:

text
=== 1. Xom feature (maydon, yil) ===
  R2 (xom): 0.986

=== 2. Feature engineering ===
  yangi: yosh, narx_kutish

=== 3. Boyitilgan feature bilan ===
  R2 (yosh bilan): 0.986

=== 4. Xulosa ===
  'yil' o'rniga 'yosh' — ma'noliroq feature
  feature engineering — model tanlovidan muhim
  ⭐ yaxshi feature — model aniqligini oshiradi

Nima ko'rsatdi: 2.1–2.8-bo'limlar.


5. To'g'ri va noto'g'ri tushunishlar

Noto'g'ri fikr To'g'risi
"Model o'zi hammani topadi" Feature yordam beradi
"Ko'p feature — yaxshi" Kam, yaxshi feature afzal
"Xom ma'lumot yetarli" Feature engineering muhim
"Sana o'zi feature" Qismlari (oy, hafta)
"Matn to'g'ridan modelga" Songa (CountVectorizer)
"Feature — texnik ish" Soha bilimi kerak
"Model tanlovi asosiy" Feature ko'pincha muhimroq
"Polinom doim yaxshi" Overfitting xavfi

6. Keng tarqalgan xatolar va yechimlari

1. Xom ma'lumot bilan cheklanish

python
model.fit(xom_X, y)   # feature'siz                # ⚠️
# hosila, sana feature qo'sh                        # ✅

2. Sanani to'g'ridan ishlatish

python
model.fit(df[["sana"]], y)   # raqam (ma'nosiz)     # ⚠️
df["oy"], df["hafta_kuni"]   # qismlar               # ✅

3. Matn to'g'ridan modelga

python
model.fit(matn, y)   # xato (matn)                  # ⚠️
CountVectorizer().fit_transform(matn)                # ✅

4. Ko'p feature (overfitting)

python
PolynomialFeatures(degree=5)   # ko'p feature       # ⚠️
# soddaroq, feature tanlov                           # ✅

5. Ma'lumot sizishi feature'da

python
df["ort"] = df["narx"].mean()   # butun ma'lumot    # ⚠️ (train'dan)
# train'dan hisobla                                  # ✅

6. Soha bilimisiz feature

python
# tasodifiy feature (ma'nosiz)                       # ⚠️
# soha bilimi (metr narxi muhim)                     # ✅

7. Binningda ma'lumot yo'qotish

python
pd.cut(yosh, bins=[0, 100])   # 1 guruh (foydasiz)  # ⚠️
# ma'noli chegaralar                                 # ✅

7. Integratsiya — bu bilim qayerda kerak bo'ladi

  • 25.3-dars (o'tilgan): Tayyorlash — feature bilan birga
  • 24.11-dars (o'tilgan): Vaqt qatorlari — sana feature
  • 25.6-dars: Ansambl — feature muhimligi
  • 25.8-dars: PyTorch — avtomatik feature
  • 25.10-dars: LLM — matn embedding

8. Eng yaxshi amaliyotlar

  1. Hosila feature (ma'noli kombinatsiya).

  2. Sana → qismlar (oy, hafta kuni).

  3. Matn → son (CountVectorizer).

  4. Kam, yaxshi feature (overfitting'dan qoch).

  5. Soha bilimini ishlat (nima muhim).

  6. Feature train'dan (ma'lumot sizishi yo'q).

  7. Feature tanlov (muhimlarini).

  8. Feature engineering'ga vaqt sarfla.


9. Amaliy topshiriq

Vazifa 1: Bashorat qiling

python
1.  # feature engineering nima?
2.  # hosila feature nima?
3.  # nisbat feature nima?
4.  # binning nima?
5.  # sana feature nima?
6.  # matn feature nima?
7.  # CountVectorizer nima?
8.  # polinom feature nima?
9.  # feature tanlovi nima?
10. # nega kam feature?
11. # soha bilimi nega kerak?
12. # feature vs model tanlovi?
Javoblar
  1. Yangi foydali xususiyatlar yaratish
  2. Mavjud ustunlardan matematik yangi
  3. Ustunlar nisbati (narx/maydon)
  4. Uzluksiz son → toifa
  5. Sanadan qismlar (oy, hafta)
  6. Matnni songa
  7. So'z → son (chastota)
  8. Daraja va ko'paytma (x^2, a×b)
  9. Eng muhim feature'larni tanlash
  10. Overfitting, shovqin
  11. Nima muhim ekanini bilish
  12. Feature ko'pincha muhimroq

Vazifa 2: Xatolarni tuzating

python
1.  model.fit(df[["sana"]], y)   # raqam       # oy, hafta

2.  model.fit(matn, y)           # matn         # CountVectorizer

3.  PolynomialFeatures(degree=5)                # soddaroq

4.  df["ort"]=df["narx"].mean()  # butun        # train'dan

5.  model.fit(xom_X, y)          # feature'siz  # hosila qo'sh
Javoblar
python
1.  df["oy"]=df["sana"].dt.month; ...

2.  CountVectorizer().fit_transform(matn)

3.  PolynomialFeatures(degree=2)  # soddaroq

4.  # train'dan hisobla

5.  df["jami"]=df["narx"]*df["soni"]; ...

Vazifa 3: Hosila

Feature:

  1. Ko'paytma
  2. Nisbat
  3. Logarifm
  4. Ma'no

Vazifa 4: Binning/sana

Feature:

  1. pd.cut
  2. dt.month
  3. dt.dayofweek
  4. Dam olish

Vazifa 5: Matn/polinom

Feature:

  1. CountVectorizer
  2. PolynomialFeatures
  3. Nega kerak
  4. Ehtiyot

Vazifa 6: Amaliy

Ta'sir:

  1. Xom feature
  2. Feature yarat
  3. Boyitilgan
  4. Solishtir

Vazifa 7: O'ylash

Feature engineering — ML ning ijodiy va ko'pincha eng ta'sirli qismi: "yaxshi feature + oddiy model > yomon feature + murakkab model". Bu soha bilimi talab qiladi (uy narxida metr narxi muhim). Nima uchun "ma'lumotni tushunish — model tanlovidan muhimroq", va nega feature engineering inson bilimi va avtomatik o'rganish (25.8 chuqur o'qitish) orasidagi farqni ko'rsatadi — klassik ML va zamonaviy AI qayerda ajraladi?

Javob

Qisqa javob: "Ma'lumotni tushunish — model tanlovidan muhimroq", chunki model ma'lumotdagi naqshni topadi, lekin yaxshi feature naqshni oson ko'rsatadi (model kuchsiz bo'lsa ham topadi). Yomon feature (xom, ma'nosiz) — hatto murakkab model ham topolmaydi. Feature engineering soha bilimi talab qiladi: uy narxida "metr narxi" (narx/maydon) muhim — buni inson biladi (soha tajribasi), model o'zi topmasa ham. Bu klassik ML va zamonaviy AI (chuqur o'qitish, 25.8) orasidagi asosiy farq: klassik ML — inson feature yaratadi (feature engineering), chuqur o'qitish — model feature'ni avtomatik o'rganadi (rasm, matn — inson tasvirlab bo'lmaydigan feature). Klassik ML'da inson bilimi hal qiladi; chuqur o'qitishda ko'p ma'lumot + hisoblash. Har ikkisi o'z joyida (kichik/tuzilgan ma'lumot → klassik + feature; katta/xom → chuqur).

1. Ma'lumot tushunish muhimroq

Model naqshni topadi, feature naqshni oson ko'rsatadi. Yaxshi feature — kuchsiz model ham topadi. Yomon feature — murakkab model ham topolmaydi.

2. Soha bilimi

Feature engineering nima muhim ekanini bilish: uy narxida metr narxi (narx/maydon), tibbiyotda simptomlar nisbati. Inson tajribasi (model o'zi bilmaydi).

3. Klassik ML vs chuqur o'qitish

Klassik ML Chuqur o'qitish (25.8)
Inson feature yaratadi Model avtomatik o'rganadi
Feature engineering Ko'p qatlam
Kichik/tuzilgan ma'lumot Katta/xom (rasm, matn)
Soha bilimi Ko'p ma'lumot + hisoblash

4. Qaysi qachon

  • Tuzilgan ma'lumot (jadval), kam → klassik + feature
  • Xom ma'lumot (rasm, matn), ko'p → chuqur o'qitish
  • Chuqur o'qitish feature'ni avtomatik (inson tasvirlab bo'lmaydi)

5. Muhandislik saboqlari

  1. Ma'lumot tushunish — model'dan muhim
  2. Feature — soha bilimi
  3. Klassik ML — inson feature
  4. Chuqur o'qitish — avtomatik feature

6. Xulosa

  1. Yaxshi feature — yarim model
  2. Soha bilimi hal qiladi
  3. Klassik (inson feature) vs chuqur (avtomatik)
  4. Har ikkisi o'z joyida

Nimani mustahkamlaydi: 2.1–2.8-bo'limlar.


Xulosa

Bu darsda feature engineering'ni o'rgandik.

Eng muhim uch fikr:

  1. Feature engineering — foydali xususiyatlar yaratish. Feature engineering — mavjud ma'lumotdan model uchun foydaliroq xususiyatlar yaratish. Hosila feature (matematik): ko'paytma (narx × soni = jami), nisbat (narx / maydon = metr narxi — ko'pincha muhim), ayirma (daromad - xarajat), logarifm (log(narx) — katta oraliqni siqadi). Maqsad: modelga ma'lumotdagi naqshni oson ko'rsatish. "Yaxshi feature — yarim model" — ko'pincha model tanlovidan muhimroq. Soha bilimi kerak.

  2. Binning, sana, matn, polinom. Guruhlash (binning — pd.cut): uzluksiz son → toifa (yosh → guruh; ba'zan toifa son'dan ma'noliroq). Sana feature 24.11-bob: dt.month (mavsumiylik), dt.dayofweek (hafta kuni), dam_olish (bool), yosh (hozir - sana). Matn feature: CountVectorizer (so'z → chastota — "bag of words"), TfidfVectorizer (muhimlik bilan) — model son kutadi. Polinom feature (PolynomialFeatures): [a, b] → [a, b, a^2, a×b, b^2] — chiziqli model chiziqsiz naqshni o'rganadi (ehtiyot: overfitting).

  3. Feature tanlovi va ijod. Feature tanlovi — eng muhim featurelarni tanlash (hammani emas): ko'p feature — overfitting 25.2-bob, shovqin. "Kam, lekin yaxshi feature" afzal. Feature engineering — ML ning ijodiy va ko'pincha eng ta'sirli qismi: yaxshi feature + oddiy model > yomon feature + murakkab model. Soha bilimi hal qiladi (nima muhim). Bu klassik ML (inson feature yaratadi) va zamonaviy chuqur o'qitish (25.8 — model feature'ni avtomatik o'rganadi — rasm, matn) orasidagi asosiy farq. "Ma'lumotni tushunish — model tanlovidan muhimroq". Klassik ML'da inson bilimi, chuqur o'qitishda ko'p ma'lumot + hisoblash (har ikkisi o'z joyida).

Keyingi darsda ansambl modellar ni o'rganamiz: ko'p modelni birlashtirib kuchliroq model (Random Forest, Gradient Boosting) — "ko'pchilik aqli" ML'da.

ML va AI · 5-dars yakunlandi

25.5-dars: Feature engineering

16 daqvaqt
0belgi
Keyingi dars: 25.6-dars: Ansambl modellar
Telegram'da
Darsda xato koʻrdingizmi?

Izohlar · 0