Teaching with and without a teacher

1) Why and when

Supervised: there is a label → we predict the probability/class/value. We use it when the "correct answer" is clear and there is a story: churn, deposit of 7 days, RG/AML risk, probability of response to an offer, LTV forecast.
Unsupervised: there are no marks → we find structures/clusters/anomalies/latent factors: segmentation of players, fraud rings, thematic profiles of games, detection of provider failures, compression of signs.

Selection rule: if the business decision depends on a specific probabilistic forecast → supervised; if the goal is to open unknown patterns/signals or to reduce the dimension of the data → unsupervised. In practice, they are combined.

2) Typical iGaming cases

Supervised

Churn/reactivation: binary classification (go/not go), uplift models for impact.
Propensity to deposit/purchase: probability of event in horizon T.
RG/AML: risk rate, structuring probability, suspicious session.
Bonus anti-abuse: the likelihood of fraudulent use of promo.
Recommendations (ranking): probability of click/bet on the game (listwise/pointwise).

Unsupervised

Player segmentation: k-means, GMM, HDBSCAN by RFM/behavior/genre.
Anomalies: Isolation Forest, LOF, AutoEncoder on payments/game patterns.
Graph analysis: clustering in the "player-device-card-IP" column.
Downsize: PCA/UMAP for visualization and feature engineering.
Thematic models: NMF/LDL for game descriptions/support chats.

3) Data and features

Point-in-time connections to exclude data leakage.
Characteristic windows: 10 min/1 h/1 day/7 days/30 days (recency, frequency, monetary).
Context: market/jurisdiction/DST/holidays, provider/genre, device/ASN.
Graph features: the number of unique cards/IP/devices, centrality.
Currency/time zone normalization, SCD II for users/games/providers.

4) Algorithms and metrics

With the teacher

Algorithms: LogReg, XGBoost/LightGBM/CatBoost, TabNet; for ranking - LambdaMART/GBDT; time series - Prophet/ETS/Gradient Boosted TS.
Metrics: ROC-AUC/PR-AUC, F1 @ operational threshold, KS (risk), NDCG/MAP @ K (recommendations), MAPE/WAPE (projections), expected cost with FP/FN weights.

Without a teacher

Clustering: k-means/GMM (number of clusters - elbow/silhouette), HDBSCAN (density).
Anomalies: Isolation Forest/LOF/AutoEncoder; metrics - precision @ k on expert markup, AUCPR on synthetic anomalies.
Dimension: PCA/UMAP for feature design and visualizations.

5) Combined approaches

Semi-Supervised: pseudo-bubbles for the part of unallocated data (self-training), consistency regulation.
Self-Supervised: contrasting/masked tasks (session/game embeddings) → use downstream in supervised.
Active Learning: the system offers marking candidates (maximum uncertainty/diversity) → saves the work of AML/RG experts.
Weak Supervision: heuristics/rules/distant markup form "weak" labels, then calibrate.

6) Process: from offline to online surfing

1. Offline: collecting/preparing → split by time/markets → training/validation → backtest.
2. Metrics semantics: uniform formulas (for example, churn_30d) and fixed time windows.
3. Feature Store: uniform feature formulas online/offline; compliance tests.
4. Online surfing: gRPC/REST endpoints, SLA by latency, AB routing/canary releases.
5. Monitoring: data/prediction drift (PSI/KL), latency p95, business metrics error, alerts.

7) Privacy and compliance

PII minimization: pseudonymization, mapping isolation, CLS/RLS.
Residency: individual pipelines/encryption keys by region (EEA/UK/BR).
DSAR/RTBF: delete/edit features and logs; keep the legal grounds for the exceptions.
Legal Hold: Freezing Investigative/Reporting Artifacts.
Fairness: Audit Proxy Feature, Impact Reports (SHAP), RG Intervention Policy.

8) Economics and productivity

The cost of calculating the feature (cost/feature) and inference (cost/request).
Materialization of offline aggregates; online - only critical windows.
Cache of permissions/scoring results for short TTL, asynchronous lookups with timeouts.
Quotas and budgets for replays/backtests; chargeback by command/model.

9) Examples (fragments)

9. 1 Point-in-time selection for churn_30d

sql
WITH base AS (
SELECT user_pseudo_id, DATE(event_time) AS asof
FROM silver. fact_events
GROUP BY user_pseudo_id, DATE(event_time)
),
feat AS (
SELECT b. user_pseudo_id, b. asof,
SUM(CASE WHEN e. type='deposit' AND e. event_time>=b. asof - INTERVAL '30' DAY
AND e. event_time<b. asof THEN amount_base ELSE 0 END) AS dep_30d,
COUNT(CASE WHEN e. type='bet' AND e. event_time>=b. asof - INTERVAL '7' DAY
AND e. event_time<b. asof THEN 1 END) AS bets_7d
FROM base b
JOIN silver. fact_events e USING (user_pseudo_id)
GROUP BY b. user_pseudo_id, b. asof
),
label AS (
SELECT f. user_pseudo_id, f. asof,
CASE WHEN NOT EXISTS (
SELECT 1 FROM silver. fact_events x
WHERE x.user_pseudo_id=f. user_pseudo_id
AND x.event_time>f. asof AND x.event_time<=f. asof + INTERVAL '30' DAY
) THEN 1 ELSE 0 END AS churn_30d
FROM feat f
)
SELECT FROM feat JOIN label USING (user_pseudo_id, asof);

9. 2 Payment anomalies (pseudocode, Isolation Forest)

python
X = build_features (payments_last_7d) # sum/frequency/novelty/BIN/ASN/time model = IsolationForest (contamination = 0. 01). fit(X_train)
scores = -model. decision_function(X_test)
alerts = where (scores> THRESHOLD) # AML case candidates

9. 3 Segmentation of k-means (RFM + genres)

python
X = scale(np. c_[R, F, M, share_slots, share_live, share_sports])
km = KMeans(n_clusters=8, n_init=20, random_state=42). fit(X)
segments = km. labels_

9. 4 Cost threshold for binary model

python threshold = pick_by_expected_cost(scores, labels, cost_fp=5. 0, cost_fn=50. 0)

10) Evaluation, validation and experiments

Offline: temporal split (train/val/test by time/markets), backtesting, bootstrap trust.
Online: A/B/n, sequential tests, CUPED/diff-in-diff.
Off-policy: IPS/DR for personalization policies.
Calibration: Platt/Isotonic for correct probabilities.
Degradation control: alerts by business metrics and PR-AUC/KS.

11) RACI

R (Responsible): Data Science (models/experiments), MLOps (platform/serving), Data Eng (features/pipelines).
A (Accountable): Head of Data/CDO.
C (Consulted): Compliance/DPO (PII/RG/AML), Security (KMS/secrets), SRE (SLO/value), Finance (ROI).
I (Informed): Product/Marketing/Operations/Support.

12) Implementation Roadmap

1. Catalog of targets/labels and signals (churn_30d, propensity_7d, risk_rg).

2. Feature Store v1 (5-10 features), basic XGBoost models, offline metrics dashboards.

3. Segmentation of k-means (8 clusters) + description of segments; Isolation Forest for payments.

4. Online surfing with cache, p95 <150 ms; A/B for 10-20% of traffic.

• Active/Semi-Supervised for Label Scarcity (AML/RG), self-supervised game/session embeddings.
• Canary releases, drift monitoring, auto retraining.
• A single semantic layer of metrics and online/offline matching feature.

• Graph signs and fraud rings; uplift bonus models.
• Multi-regional serving, quotas/chargeback; WORM archive of releases.
• Fairness audit, stress tests, runbooks incidents.

13) Pre-sale checklist

• Point-in-time sampling and anti-leakage tests.
• Probability calibration; Select the expected cost threshold.
• Model cards (owner, data, metrics, risks, fairness).
• Feature Store Online/Offline Compliance Test.
• Drift/latency/error monitoring, alerts and auto-rollback.
• PII/DSAR/RTBF/Legal Hold policies; logging is impersonal.
• Plan A/B and statistical power calculated; The rollback runbook is ready.

14) Anti-patterns

Mixing new events into labels (leakage) and the absence of point-in-time.
"One model for all" instead of domain decomposition.
Some librated probabilities → incorrect business thresholds.
Blind flight: no online drift/quality monitoring.
Online overcomplication (heavy external-joins without cache and timeouts).
Segments without business interpretation and owner.

15) The bottom line

Supervised learning provides measurable prognosis and risk/income management; without a teacher - structure and signals where there are no marks. Their combination (semi/self-supervised, active learning) in data discipline (point-in-time, Feature Store), compliance and MLOps gives the iGaming platform a steady increase in Net Revenue, a decrease in fraud and timely RG interventions - with reproducibility, cost control and readiness for audit.