Course Objective & Scope
When user behaviour, market conditions or fraud patterns shift within minutes, weekly batch retraining lags behind reality. This course covers online machine learning algorithms that learn instance by instance or in micro-batches (online logistic regression, Hoeffding Trees, Adaptive Random Forests), the River library, streaming feature statistics and test-then-train (prequential) evaluation.
Real-time concept drift detection with ADWIN, DDM and Page-Hinkley, handling delayed labels and contextual bandits for immediate-feedback decisions are practised hands-on. Participants benchmark a static batch baseline against an online/hybrid updating model on a drifting data stream and design a safe production architecture.
Who Should Attend?
- Data scientists and ML engineers building real-time scoring and prediction systems
- Fraud, dynamic pricing, recommendation and IoT telemetry teams
- Platform engineers running analytics on Kafka/Flink event streams
- Prerequisite: Python and core machine learning experience
Key Takeaways
- Decide when frequent batch retraining suffices vs when true online learning is warranted
- Build streaming preprocessing, rolling features and incremental model updates with River
- Implement prequential (test-then-train) evaluation and handle delayed feedback correctly
- Detect concept drift with ADWIN and Page-Hinkley to trigger controlled model adaptation
- Guard live-updating models against catastrophic forgetting and corrupted stream bursts
Curriculum
Batch vs Online Learning and the Streaming Paradigm
- Bounded memory and single-pass constraints
- Streaming standardisation, target encoding and rolling stats in River
Incremental Learning Algorithms
- SGD linear models and Passive-Aggressive learners
- Hoeffding Trees, Hoeffding Adaptive Trees and Adaptive Random Forests
Prequential Evaluation and Delayed Labels
- Test-then-train loops and rolling-window metrics
- Verification latency and importance weighting
Concept Drift Detection and Adaptation
- Sudden, gradual and recurring drift; ADWIN, DDM, Page-Hinkley
- Drift-triggered retraining, shadow models and contextual bandits
Production Guardrails and Hybrid Architectures
- Warm-start batch base models with an online residual/bandit layer
- Update freezing, rollback and guardrails against poisoned stream bursts
Other Courses in This Category
-
Modelling relational data with NetworkX and PyTorch Geometric: centrality and community algorithms, graph features, Node2Vec and GCN/GraphSAGE/GAT for node, link and subgraph prediction in fraud rings and supply networks.
- Live Course
- 5 modules
-
In-context learning on tabular data with TabPFN v2, plus zero-shot and fine-tuned time-series forecasting with Chronos, TimesFM, Moirai and Lag-Llama, benchmarked against GBDTs and classical models.
- Live Course
- 5 modules
-
Supervised and unsupervised learning, train/test splits, overfitting, core metrics and first models in scikit-learn, so technical teams can frame a business problem as an ML task and build a sound baseline.
- Live Course
- 5 modules
-
Gradient boosting, ensembles and stacking, hyperparameter optimisation with Optuna, imbalanced data, probability calibration and custom loss functions for experienced data scientists building production-grade models.
- Live Course
- 5 modules
-
Takes analysts beyond classic reporting with segmentation, driver analysis, anomaly flagging and simple predictive models in Python, scikit-learn and SHAP, producing analyses that tie insight to business decisions.
- Live Course
- 5 modules
-
Categorical encoding, time and window-based features, target encoding, interaction and text features, feature selection and feature stores: a systematic approach to the step that most improves model performance on tabular data.
- Live Course
- 5 modules
Get in touch
