Unsupervised — label yoki oldindan berilgan targetlarsiz data ichidagi structure, pattern, similarity va yashirin representationni o‘rganadigan machine-learning yondashuvi. To‘liq nomi Unsupervised Learning. Clustering, dimensionality reduction, density estimation va anomaly discovery uning keng vazifalaridir.
Unsupervised natija “to‘g‘ri javob” bilan bevosita o‘lchanmagani sabab interpretation va evaluation murakkabroq.
Unlabeled data
Dataset faqat input samplelardan iborat:
x1, x2, x3, ...
Class yoki numeric target yo‘q.
Bunday data yig‘ish labeled datasetdan arzonroq.
Ammo preprocessing va source sifati baribir muhim.
Clustering
O‘xshash samplelar guruhlanadi.
- K-means;
- hierarchical;
- DBSCAN;
- Gaussian mixture;
- spectral clustering.
Clusterlar keyin inson tomonidan tahlil qilinadi.
Cluster IDning o‘zi ma’noli label emas.
Dimensionality reduction
Ko‘p feature kichik representationga tushiriladi.
Usullar:
- PCA;
- ICA;
- NMF;
- autoencoder;
- t-SNE;
- UMAP.
Bu visualization, compression, noise reduction va downstream modelga yordam beradi.
PCA
Principal Component Analysis variance eng katta bo‘lgan orthogonal yo‘nalishlarni topadi.
Linear transform.
Feature scalingga sezgir.
Componentlar original featurelarning weighted kombinatsiyasi.
Katta variance har doim task uchun eng muhim signal emas.
Autoencoder
Encoder inputni kichik latent representationga aylantiradi.
Decoder inputni qayta tiklashga harakat qiladi.
Reconstruction loss orqali labeled targetsiz o‘rganadi.
Model juda kuchli bo‘lsa ma’noli compression o‘rniga identity mappingga yaqinlashishi mumkin.
Density estimation
Data distributioni qayerda zich ekanini model qiladi.
Yangi sample probability yoki density score oladi.
Gaussian mixture, kernel density va generative model ishlatilishi mumkin.
High-dimensional density estimation murakkab.
Anomaly detection
Ko‘pchilik patternidan farq qiladigan sample topiladi.
Usullar:
- Isolation Forest;
- One-Class SVM;
- autoencoder error;
- density;
- distance;
- clustering.
Rare sample avtomatik zararli yoki xato degani emas.
Representation learning
Model raw datadan downstream vazifa uchun foydali feature o‘rganadi.
Self-supervised learning unlabeled data’dan synthetic objective yaratadi.
Masalan, masklangan tokenni yoki rasmning boshqa view’ini bashorat qilish.
Bu zamonaviy pretrainingning asosidir.
Association rule
Transactionlarda birga uchraydigan itemlar topiladi.
Masalan:
A olganlar Bni ham oladi
Support, confidence va lift rule kuchini baholaydi.
Correlation causal relationship emas.
Topic modeling
Document collection ichidagi yashirin mavzularni topadi.
LDA va matrix factorizationga oid usullar so‘z-topic distribution yaratadi.
Topiclar avtomatik nomlanmaydi.
Inson eng yuqori so‘zlarga qarab ma’no beradi.
Feature scaling
Distance va variance’ga asoslangan unsupervised usullar scale’ga sezgir.
Bir feature katta diapazonda bo‘lsa clusterlarni o‘zi belgilab qo‘yishi mumkin.
Standardization, robust scaling yoki domain transform ishlatiladi.
Hyperparameter
Label yo‘qligi sabab cluster soni, epsilon va latent dimension tanlash qiyin.
Internal metric:
- silhouette;
- reconstruction;
- likelihood;
- stability.
Business usefulness va domain review ham zarur.
Evaluation
Ground-truth bo‘lmasa:
- cluster cohesion;
- separation;
- stability;
- reconstruction;
- downstream task;
- human interpretation;
- business outcome
bilan baholanadi.
Chiroyli 2D visualization yaxshi model isboti emas.
Data leakage
Unsupervised preprocessing ham test data’dan signal olishi mumkin.
Masalan, PCA yoki scaler butun datasetda fit qilinsa test distribution trainingga kiradi.
Evaluation pipeline’da transform faqat train splitda fit qilinadi.
Unsupervised va supervised
Supervised aniq labelga predictionni optimallashtiradi.
Unsupervised dataning ichki structure’ini topadi.
Unsupervised representation keyin kichik labeled dataset bilan supervised taskda ishlatilishi mumkin.
Self-supervised bilan farqi
Self-supervised usul labelni dataning o‘zidan yaratadi.
Masalan, yashirilgan token asl target.
Texnik jihatdan loss va target mavjud, ammo inson annotationi kerak emas.
U ko‘pincha unsupervised keng oilasiga yaqin tilga olinadi.
Use case
Unsupervised learning:
- customer segmentation;
- document grouping;
- fraud candidate;
- image organization;
- compression;
- exploratory analysis;
- pretraining;
- recommendation.
Natija production qaroriga aylanishidan oldin ma’nosi tekshiriladi.
Cheklovlar
Algorithm data’dagi har qanday structure’ni topishi mumkin, ammo u business uchun muhim bo‘lmasligi ehtimoli bor.
Feature bias va collection bias natijada saqlanadi.
Clusterga nom berish inson stereotipini qo‘shishi mumkin.
Sensitive guruhlar ehtiyotkor tahlil qilinadi.
Reconstruction
Autoencoder va matrix factorization inputni qayta tiklash xatosini minimallashtiradi. Past reconstruction error foydali semantic representationni kafolatlamaydi. Model mayda noise’ni ham yodlab olishi mumkin. Downstream task va human inspection bilan tekshiriladi.
Identifiability
Bir xil data bir nechta teng darajada yaxshi latent yoki cluster representationga ega bo‘lishi mumkin. Cluster ID, component sign va axis runlar orasida almashishi mumkin. Natijani absolute nomlar bilan emas, stable relationship va task foydasi bilan talqin qilish kerak.
Pretext task
Self-supervised model masklangan qismni topish, keyingi frame’ni bashorat qilish yoki ikki view’ni yaqinlashtirish kabi sun’iy vazifada o‘qitiladi. Pretext task downstream ma’lumotga mos feature o‘rgatishi kerak.
Compute
Unlabeled data ko‘p bo‘lsa preprocessing va training qimmatlashadi. Sample selection va distributed pipeline zarur bo‘lishi mumkin.
Initialization
K-means va neural latent model random boshlanishga sezgir bo‘lishi mumkin. Bir nechta seed bilan training qilib, objective va stability solishtiriladi. Faqat eng chiroyli visualization tanlanmaydi.
Bog‘liq tushunchalar
Unsupervised Learning, Clustering, Dimensionality reduction, PCA, Autoencoder, Density estimation, Anomaly detection, Topic modeling, Representation learning, Self-supervised learning