iGaming bilan mashq qilish

1) Biznes-keyslar va qiymati

Mahsulot/daromad: prognoz LTV, churn (chiqib ketish), depozit/sotib olishga propensiti, dinamik missiyalar/kvestlar, next-best-action/offer.
Marketing/CRM: look-alike, segmentatsiya, real-time triggerlar, bonuslarni optimallashtirish (ABO - Abuse-resistant Bonus Optimization).
Xavf/Komplayens: antifrod/AML (velocity, tuzilish, grafik belgilar), Responsible Gaming (RG) - xavf-skor, aralashuv triggerlari.
Operatsiyalar/SRE: hodisalarni bashorat qilish, capacity/traffic forecasting, provayderlarning anomaliyalari.
Moliya: GGR/NGR prognozi, Fx-sezgirlik, kontragentlar manipulyatsiyasining deteksiyasi.

Ta’sir ko’rsatkichlari: personallashtirish hisobiga Net Revenue ga + 3-7%, fraud-loss ga − 20-40%, churn ga − 10-25%, RG <5 s reaksiyaga kirishish SLA.

2) Ma’lumotlar va belgilar (Feature Engineering)

Manbalar: gameplay, to’lovlar/PSP, autentifikatsiya, qurilmalar/ASN/geo, RG/KYC/KYB, marketing UTM, provayder loglari, sapport/matnlar.

• Xulq-atvor oynalari: N stavkalar/depozitlar va 10 min/soat/kun uchun summalar, recency/frequency/monetary.
• Ketma-ketlik: o’yinlar zanjiri, oxirgi faollik bilan vaqt, sessiya belgilari.
• Geo/qurilma: mamlakat/bozor, ASN, qurilma/brauzer turi.
• Grafik: o’yinchi-xarita-qurilma-IP aloqalari, komponentlar/markazliklar (fraud rings).
• Kontekst: sutka vaqti/hafta kuni/bozor bayramlari, provayder/janr/o’yinning o’zgaruvchanligi.
• RG/AML: limitlar, o’z-o’zini istisno qilish, skrining bayroqlari, RER/sanksiyalar (kesh/asinxron orqali).

• Valyuta va vaqtni normallashtiring (UTC + bozor lokali).
• Oʻlchamlarni tarixlashtiring (SCD II).
• Onlayn/oflayn transformatsiyani (Feature Store-da yagona kod) muvofiqlashtiring.

3) Arxitektura: onlayn

3. 1 Oflayn kontur

Lakehouse: Bronze → Silver (normallashtirish/boyitish) → Gold (datasetlar).
Feature Store (offline): formula registri fich, point-in-time join, o’qitish tanlanmalarini materiallashtirish.
Trening: qat’iy bog’liqliklarga ega konteynyerlar; eksperimentlarning trekingi (metrika/artefaktlar/ma’lumotlar).
Validatsiya: k-fold/temporal split, backtest, off-policy baholash.

3. 2 Onlayn kontur

Ingest → Stream Processing: Flink/Spark/Beam bilan derazalar/watermarks, idempotentlik.
Feature Store (online): past patentli kesh (Redis/Scylla) + oflayn qoliplar.
Serving: REST/gRPC endpointlari, skoring grafigi, AB-routing, kanar relizlari.
Real-time vitrinalar: Panellar/qoidalar uchun ClickHouse/Pinot.

4) Namunaviy modellar va yondashuvlar

Tasniflash/skoring: churn/depozit/frod/RG (LogReg, XGBoost/LightGBM, TabNet, CatBoost).
Rang berish/tavsiyalar: faktorizatsiya/list-ranging (LambdaMART), seq2rec (RNN/Transformers), kontekstli banditlar.
Vaqtinchalik qatorlar uchun: Isolation Forest, One-Class SVM, AutoEncoder, Prophet/TSfresh.
Grafik: firibgarlik halqalari uchun Node2Vec/GraphSAGE/GNN.
Sababi (causal): uplift-modellar, T-learner/X-learner, DoWhy/CausalML.
NLP/ASR: tiketlar/chatlar, shikoyatlar tasnifi, sentiment, mavzular.

5) Sifat metrikasi

Tasniflash: ROC-AUC/PR-AUC, operatsion ostonalarda F1, expected cost (tortilgan FP/FN), risk-skoring uchun KS.
Tavsiyalar: NDCG @K, MAP @K, coverage/diversity, CTR/CVR onlayn.
TS/Forecast: MAPE/SMAPE, WAPE, P50/P90 xatosi, PI qoplamasi.
RG/AML: SLAda precision/recall, o’rtacha vaqt-to-intervene.
Iqtisodiyot: Net Revenue-da uplift, fraud saved, ROI kampaniyalari,% bonus-abyuza.

6) Baholash va eksperimentlar

Oflayn: temporal split, haftalar/bozorlar/tenantlar bo’yicha backtest.
Onlayn: A/B/n, CUPED/diff-v-diff, sequential tests.
Off-policy: Personalizatsiya siyosati uchun IPS/DR.
Stat. quvvat: dispersiya va MDE hisobga olingan holda tanlov hajmini hisoblash.

python cost_fp = 5. 0 # false alarm cost_fn = 50. 0 # missed fraud threshold = pick_by_expected_cost (scores, labels, cost_fp, cost_fn)

7) Maxfiylik, axloq, komplayens

PII-minimallashtirish: taxalluslar, mappinglarni izolyatsiya qilish, CLS/RLS.
Rezidentlik: EEA/UK/BR alohida konturlari; asossiz kross-mintaqaviy join’onlarsiz.
DSAR/RTBF: oʻchirish/fich va log tahriri; Sex/hisobot uchun Legal Hold.
Fairness/xolislik: fich auditi, disparate impact, proxy-oʻzgaruvchilarni boshqarish.
Explainability: SHAP/feature importance, model kartochkalari (owner, sana, ma’lumotlar, metriklar, xavflar).
Xavfsizlik: KMS/CMK, jurnaldan tashqari sirlar, WORM-relizlar arxivlari.

8) MLOps: hayot sikli

1. Data & Features: sxemalar/kontraktlar, DQ qoidalari (completeness/uniqueness/range/temporal), lineage.
2. Trening: konteynerlar, avtotyuning, treking tajribalari.
3. Validatsiya: sxemalar mosligi testlari, bias/fairness, performance testlari.
4. Reliz (CI/CD/CT): kanar/bosqichma-bosqich chiqish, ficha-bayroqlar, «qorong’u ishga tushirish».
5. Serving: avtoskeyling, keshlash, gRPC/REST, timeouts/retrai.
6. Monitoring: ma’lumotlar/prognozlar dreyfi (PSI/KL), latency p95, error-rate, coverage, «silent metrics».
7. Re-train: metriklarning dreyf/degradatsiyasi jadvali/triggerlari.
8. Hodisalar: runbook, qaytarish modeli, fallback (qoida/oddiy model).

9) Feature Store (kelishuv yadrosi)

Oflayn: point-in-time hisoblash, anti-leakage, fich formulasining versiyasi.
Onlayn: past latentlik (10-30 ms ≤), TTL, oflayn rejimga muvofiqlik.
Kontraktlar: nomi/tavsifi, egasi, SLA, formula, muvofiqlik testlari online/offline.

yaml name: deposits_sum_10m owner: ml-risk slo: {latency_ms_p95: 20, availability: 0. 999}
offline:
source: silver. payments transform: "SUM(amount_base) OVER 10m BY user_pseudo_id"
online:
compute: "streaming_window: 10m"
tests:
- compare_online_offline_max_abs_diff: 0. 5

10) Onlayn skoring va qoidalar

Gibrid ML + Rules: model → tezkor + tushuntirish; qoidalar - hard-guard/etika/qonun.
Tikish: CEP-patternlar (structuring/velocity/device switch) + ML-skoring.
SLA: p95 end-tu-end 50-150 ms personalizatsiya uchun, ≤ 2-5 s RG/AML alertlari uchun.

python features = feature_store. fetch(user_id)
score = model. predict(features)
if score > T_RG:
trigger_intervention(user_id, reason="RG_HIGH_RISK", score=score)
elif score > T_BONUS:
send_personal_offer(user_id, offer=choose_offer(score, seg))

11) O’qitish uchun ma’lumotlar: tanlov va leybllar

Hodisa oynalari: t0 - referens, t0 + Δ - leybl (depozit/qora/frod).
Leakage-control: point-in-time join, kelajakdagi voqealarni istisno qilish.
Balanslash: tabaqalarning tabaqalanishi/og’irligi, noyob sinflar uchun focal loss.
Etika: sezgir atributlarni/proksilarni istisno qilish, ta’sirni nazorat qilish.

12) Iqtisodiyot va unumdorlik

Fich: cost/feature va cost/request deb hisoblang, og’ir onlayn-join’lardan qoching.
Kesh: RAMda issiq chichlar, sovuq - lazy.
Materiallashtirish: oflayn agregatsiya; onlayn faqat tanqidiy.
Kvotalar: vaqt oynalari bo’yicha repleylar, bektestlar uchun limitlar; buyruqlar bo’yicha chargeback.

13) SQL/psevdokod namunalari

sql
WITH base AS (
SELECT user_pseudo_id, MIN(event_time) AS first_seen
FROM silver. fact_bets
GROUP BY user_pseudo_id
),
agg AS (
SELECT user_pseudo_id,
DATE(t. event_time) AS asof,
SUM(amount_base) FILTER (WHERE type='deposit' AND event_time >= t. event_time - INTERVAL '30' DAY AND event_time < t. event_time) AS dep_30d,
COUNT() FILTER (WHERE type='bet' AND event_time >= t. event_time - INTERVAL '7' DAY) AS bets_7d
FROM silver. fact_events t
GROUP BY user_pseudo_id, DATE(t. event_time)
)
SELECT a. user_pseudo_id, a. asof, a. dep_30d, a. bets_7d,
CASE WHEN NOT EXISTS (
SELECT 1 FROM silver. fact_events e
WHERE e. user_pseudo_id=a. user_pseudo_id AND e. event_time > a. asof AND e. event_time <= a. asof + INTERVAL '30' DAY
) THEN 1 ELSE 0 END AS label_churn_30d
FROM agg a;

sql
SELECT user_id,
TUMBLE_START(event_time, INTERVAL '10' MINUTE) AS win_start,
COUNT() AS deposits_10m,
SUM(amount_base) AS sum_10m
FROM stream. payments
GROUP BY user_id, TUMBLE(event_time, INTERVAL '10' MINUTE);

14) Joriy etish yo’l xaritasi

1. Signallar katalogi va Feature Store v1 (Payments/Gameplay uchun 5-10 fich).

2. Asosiy model churn/depozit (XGBoost) + A/B 10-20% trafikka.

3. Kesh (p95 <150 ms) va kanareya relizlari bilan onlayn serving.

4. Dreyf/sifat monitoringi, model kartochkasi, qaytarish runbook.

• RG/AML-skoringlar, grafik belgilar, real-time triggerlar.
• Bonuslar uchun Uplift modellari, kontekstli banditlar, off-policy baholash.
• Dreyf/kalendar bo’yicha avto-re-treyn, hujjatlarni avtomatlashtirish.

• O’yinlar katalogini (seq2rec) personallashtirish, ko’p obyektiv optimallashtirish (daromad/javobgarlik).
• Ko’p mintaqaviy serving, SLAs/kvotalar, chi/inferens bo’yicha chargeback.
• Fairness-audit va stress-testlar, DR-mashqlar va WORM-relizlar omborlari.

15) RACI

R (Responsible): MLOps (platforma/serving), Data Science (modellar/eksperimentlar), Data Eng (fich/payplaynlar).
A (Accountable): Head of Data / CDO.
C (Consulted): Compliance/DPO (PII/RG/AML/DSAR), Security (KMS/sirlar), SRE (SLO/qiymat), Finance (effekt/ROI), Legal.
I (Informed): Mahsulot/Marketing/Operatsiyalar/Qo’llab-quvvatlash.

16) Sotishdan oldingi chek-varaq

• Chichlar onlayn/offline orqali kelishilgan, reproduktiv testlar o’tkazilgan.
• Model kartochkasi (owner, ma’lumotlar, metriklar, xatarlar, fairness) to’ldirilgan.
• Kanar relizi/fichflag; SLA va latency/xato/dreyf alertlari.
• PII/DSAR/RTBF/Legal Hold siyosatiga rioya qilingan; loglar nomaʼlum.
• Hodisalar/orqaga qaytish Runbook; fallback strategiyasi.
• Eksperimentlar rasmiylashtirilgan (gipotezalar, metriklar, davomiyligi, MDE).
• Inferens va fich qiymati budjetga yozilgan; kvotalar va limitlar kiritilgan.

17) Anti-patternlar

Tafovut onlayn/oflayn → buzilmaslik.
Kesh va taymautsiz «issiq yo’lda» sinxron tashqi API.
Noaniq metrik formulalar/model kartochkalari yo’qligi.
Monitoring va qayta mashq qilmasdan qayta o’qitish/dreyf.
CLS/RLS/minimallashtirmasdan tahlil va mashg’ulotlarda PII.
«Hamma uchun bitta katta model» domen dekompozitsiyasisiz.

18) Jami

ML iGaming - «sehrli» modellar to’plami emas, balki intizom: muvofiqlashtirilgan ma’lumotlar va fichlar, takrorlanadigan oflayn-trening, ishonchli onlayn serving, qat’iy MLOps, shaffof metrika va axloq/komplayens. Ushbu yoʻl-yoʻriqqa amal qilib, siz daromad va ushlab qolishni barqaror oshiradigan, tavakkalchiliklarni kamaytiradigan va tartibga solish talablariga rioya qiladigan tizim qurasiz - masshtabda, tez va oldindan aytish mumkin.