Classification — input sample’ni oldindan belgilangan class yoki categorylardan biriga yoki bir nechtasiga ajratadigan supervised machine-learning vazifasi. Emailni spam deb aniqlash, rasm objectini tanish, kasallik xavfini categorylash va support murojaatini mavzuga ajratish classification misollaridir.
Model trainingda labeled datasetdan classlar orasidagi qaror chegarasini o‘rganadi.
Binary classification
Ikki class mavjud:
positive / negative
fraud / normal
spam / not spam
Model bitta score yoki probability chiqarishi mumkin.
Thresholddan yuqori bo‘lsa positive deb belgilanadi.
Threshold 0.5 bo‘lishi shart emas.
Multi-class classification
Har sample faqat bitta classga tegishli:
cat | dog | bird
Model har class uchun logit chiqaradi.
Softmax ularni probability distributionga aylantiradi.
Yakuniy class eng katta probability bilan tanlanishi mumkin.
Multi-label classification
Bir sample bir nechta classga ega:
["security", "network", "linux"]
Har class odatda mustaqil sigmoid outputga ega.
Threshold har class uchun boshqacha bo‘lishi mumkin.
Classlar o‘zaro bog‘liq bo‘lsa model bu relationshipni ham o‘rganishi mumkin.
Logit
Modelning softmax yoki sigmoiddan oldingi raw outputi logit.
Logit cheklanmagan son.
Loss ko‘pincha numerical stability uchun probabilityga aylantirmasdan logits bilan ishlaydi.
Inference’da probability yoki score sifatida transform qilinadi.
Cross-entropy
Classificationda keng ishlatiladigan loss.
To‘g‘ri class probabilitysi past bo‘lsa katta xato beradi.
Multi-class uchun categorical cross-entropy, binary va multi-label uchun binary cross-entropy ishlatiladi.
Class imbalance weighting bilan boshqarilishi mumkin.
Decision boundary
Feature space’ni classlarga ajratadigan chegara.
Linear classifier tekis hyperplane yaratadi.
Tree, kernel va neural network murakkab nonlinear boundary o‘rganishi mumkin.
Juda murakkab boundary training data’ga overfit qilishi ehtimoli bor.
Logistic regression
Nomida regression bo‘lsa ham binary classification algoritmi.
Linear score sigmoid orqali probabilityga aylanadi.
Weightlar feature ta’sirini ko‘rsatishi mumkin.
Regularization ko‘p featureda overfittingni kamaytiradi.
Decision tree
Feature qiymatlariga qarab branchlar yaratadi.
Leaf class yoki probability saqlaydi.
Interpretatsiya qulay.
Chuqur tree training data’ni yodlab qolishi mumkin.
Random Forest va boosting tree’larni ensemble qiladi.
Support Vector Machine
SVM classlar orasidagi marginni maksimal qilishga intiladi.
Kernel nonlinear boundary yaratishi mumkin.
Kichik va o‘rta datasetda kuchli.
Juda katta dataset va probability calibrationda qo‘shimcha xarajat mavjud.
Neural classifier
Neural network feature yoki raw inputdan representation o‘rganadi.
Output head class logitsini beradi.
Image, text va audio classificationda pretrained backbone keng ishlatiladi.
Fine-tuning dataset va class definitionga mos bo‘ladi.
Class imbalance
Kam class muhim bo‘lishi mumkin.
Masalan, fraud 0.5%.
Accuracy model foydasini ko‘rsatmaydi.
Usullar:
- class weight;
- oversampling;
- undersampling;
- focal loss;
- threshold tuning;
- anomaly approach.
Confusion matrix
Haqiqiy va predicted class kombinatsiyasini ko‘rsatadi.
Binary holatda:
- true positive;
- false positive;
- true negative;
- false negative.
Har xato turi business xarajatga ega.
Precision
Positive deb topilganlarning qanchasi haqiqiy positive:
TP / (TP + FP)
False positive qimmat bo‘lsa precision muhim.
Spam filter qonuniy emailni bloklamasligi kerak.
Recall
Haqiqiy positivelarning qanchasi topilgan:
TP / (TP + FN)
False negative qimmat bo‘lsa recall muhim.
Kasallik screeningda xavfli holatni o‘tkazib yubormaslik kerak.
F1 score
Precision va recallning harmonic mean’i.
Ikkisini bitta metricda muvozanatlashtiradi.
True negative ko‘pligini hisobga olmaydi.
Class imbalance’da accuracydan foydaliroq bo‘lishi mumkin.
ROC va PR curve
Threshold o‘zgarganda metriclar qanday almashishini ko‘rsatadi.
ROC true-positive va false-positive rate’ni chizadi.
Precision-recall curve kam positive classda ko‘proq ma’noli bo‘lishi mumkin.
AUC ranking sifatini umumlashtiradi.
Threshold
Probabilitydan classga o‘tish chegarasi.
Business xarajat va desired precision-recall asosida tanlanadi.
Production data distributioni o‘zgarsa threshold qayta kalibrlanadi.
Har subgroup uchun boshqa threshold fairness masalasini keltirishi mumkin.
Calibration
0.8 probability berilgan samplelarning taxminan 80%i to‘g‘ri bo‘lishi kerak degan ma’no calibrationga bog‘liq.
Model rankingda yaxshi, probabilityda yomon bo‘lishi mumkin.
Platt scaling va isotonic regression kabi usullar qo‘llanadi.
Open-set muammo
Productionda trainingda bo‘lmagan yangi class kelishi mumkin.
Oddiy classifier baribir ma’lum classlardan birini tanlaydi.
Unknown detection, confidence threshold yoki embedding distance ishlatilishi mumkin.
Hierarchical classification
Classlar daraxt ko‘rinishida bo‘lishi mumkin:
Texnologiya
→ Dasturlash
→ Python
Model avval yuqori kategoriya, keyin pastki classni tanlashi yoki barcha darajani birgalikda bashorat qilishi mumkin. Parent va child predictionlari o‘zaro mos bo‘lishi kerak.
Cost-sensitive classification
False positive va false negative xarajati teng emas. Fraudni o‘tkazib yuborish katta zarar, qonuniy transactionni bloklash esa user tajribasiga zarar yetkazadi. Loss weight va threshold business cost matrix asosida tanlanadi.
Abstention
Model confidence past bo‘lsa class berish o‘rniga “aniq emas” deb human reviewga yuborishi mumkin. Coverage kamayadi, ammo qolgan predictionlar aniqligi oshadi. Threshold calibration bilan belgilanadi.
Hierarchical metric
Parent category to‘g‘ri, child category xato bo‘lsa bu mutlaqo boshqa branchdagi xatodan yengilroq bo‘lishi mumkin. Hierarchical precision yoki tree distance shunday farqni o‘lchaydi.
Bog‘liq tushunchalar
Classification, Binary classification, Multi-class classification, Multi-label classification, Cross-entropy, Precision, Recall, F1 score, Confusion matrix, Calibration