Course Description
In the age of data, the key to gaining a competitive edge is extracting actionable insights from data.
This comprehensive training offers participants a practical journey starting from the fundamentals of machine learning all the way to solving complex real-world problems.
Through hands-on practice with Python and the scikit-learn ecosystem, you will learn data preprocessing, model building, hyperparameter optimization, and model evaluation workflows.
You will be able to build predictive models using modern algorithms such as XGBoost and LightGBM, explain your models with SHAP, and effectively handle imbalanced datasets.
Target Audience
This training is designed for professionals who want to take an active role in data-driven decision-making and create business value with machine learning.
Data analysts and data scientists: Ideal for those looking to deepen their ML skills and master modern algorithms.
Software developers: Beneficial for engineers aiming to add prediction and classification capabilities to their applications.
Business analysts and product managers: Gain the technical foundation to develop data-driven strategies and manage ML projects effectively.
Engineers, researchers, and academics: Build a solid foundation for applying ML in production optimization, quality control, or scientific research.
Professionals in finance, marketing, sales, and operations: Anyone looking to improve business processes with predictive modeling.
Prerequisites
Basic knowledge of Python programming (variables, loops, functions)
Familiarity with core statistical concepts (mean, standard deviation, correlation)
Beginner-level experience with the Pandas library (preferred, not required)
Learning Outcomes
Understand the fundamentals of machine learning and distinguish between different ML learning paradigms
Apply train/validation/test split principles and detect and resolve overfitting and underfitting
Select and implement appropriate cross-validation strategies (K-Fold, StratifiedKFold, TimeSeriesSplit)
Detect and handle missing data and outliers to prepare datasets for modeling
Engineer new features and transform categorical variables using modern encoding techniques
Build and optimize modern ML algorithms (Linear/Logistic Regression, Decision Trees, Random Forest, XGBoost, LightGBM, CatBoost)
Perform hyperparameter optimization with Optuna and choose between Grid Search, Random Search, and Bayesian Optimization
Build high-performing models on imbalanced datasets using SMOTE, class weights, and threshold optimization
Provide model explainability (SHAP, LIME, Permutation Importance) and communicate insights to business stakeholders
Design end-to-end ML workflows with Scikit-learn Pipeline and ColumnTransformer while preventing data leakage
Analyze and interpret model performance using confusion matrices, ROC-AUC, PR-AUC, and learning curves
Curriculum
Module 1 - Machine Learning Fundamentals
Introduction to Machine Learning (ML)
What Is Machine Learning?
Definition of ML
How it differs from traditional programming
Real-world applications and examples
When to use ML and when not to
Types of Machine Learning
Supervised vs. Unsupervised Learning
Classification vs. Regression
Training vs. Prediction phases
Core ML Concepts
Features and targets
The golden rule: Never touch the test data before final evaluation
Overfitting and underfitting
Why validation is critical
Evaluation metric vs. loss function
How to choose an error/loss function
The ML Lifecycle
Data Collection → Preprocessing → Feature Engineering → Modeling → Evaluation → Deployment
Iterative nature of ML projects
Common pitfalls and how to avoid them
Module 2 - Data Splitting and Preprocessing
Data Splitting
Three-Way Split
Training set
Validation set
Test set - kept locked away
Why a simple train/test split alone is dangerous
Introduction to Cross-Validation
What is K-fold cross-validation?
Choosing the right K value
Computational cost trade-offs
Visualizing splits
When not to use standard K-fold
Advanced Validation Strategies
TimeSeriesSplit
GroupKFold
StratifiedKFold
Multi-label StratifiedKFold with iterative stratification
Nested Cross-Validation
Data Preprocessing Fundamentals
Feature Engineering Basics
Creating new features
Domain-specific feature analysis
Creating interaction features from existing variables
Feature scaling and normalization
Why preprocessing matters
Handling Categorical Variables
Label encoding
One-hot encoding
Target encoding fundamentals
Advanced encoding strategies
When to use each encoding method
Feature Selection
Removing irrelevant features
Correlation analysis
Feature importance fundamentals
Data Preprocessing - Outliers
Outlier Detection
Statistical methods: IQR, Z-score, Modified Z-score
ML-based methods: Isolation Forest, Local Outlier Factor, DBSCAN
Domain-specific threshold rules
Multivariate outlier detection
Visual analysis: box plots, scatter plots, Q-Q plots
Handling Outliers - Removal
When to remove: true errors and data entry mistakes
Percentage threshold rules (e.g., investigate if >5%)
Impact analysis before removal
Handling Outliers - Capping / Winsorization
Percentile capping (e.g., 1st and 99th percentiles)
Applying Winsorization
Reducing outlier impact while preserving dataset size
Handling Outliers - Transformations
Log transformation for right-skewed data
Square root and cube root transformations
Box-Cox transformation for optimal normality
RobustScaler to minimize outlier influence
Handling Outliers - Binning
Discretizing continuous values into categorical bins
Equal-width vs. equal-frequency binning
Custom bins based on domain knowledge
Handling Outliers - Other Methods
Building separate models for outliers
Outlier flagging as a binary feature
Ensemble approach: normal + outlier models
Treating outliers as missing values and imputing
Using robust statistics (median, trimmed mean)
Model-based imputation for outliers
Why different model families require different outlier handling strategies
Data Preprocessing - Missing Data
Handling Missing Values
Dropping missing data: pitfalls and when it is acceptable
Simple imputation strategies and their limitations
Advanced imputation: IterativeImputer, KNNImputer, MICE
When to drop vs. when to impute
Creating missingness indicators
Domain-specific imputation strategies
Incorporating missingness as an informative feature
Module 3 - Building ML Models
Pipeline Architecture and Baseline Models
Building Pipelines
Scikit-learn Pipeline and ColumnTransformer
Preventing data leakage through proper pipeline design
Creating custom transformers
Pipeline persistence and versioning
Establishing a Baseline Model
The baseline modeling mindset
Simple linear/logistic regression baseline
AutoML baselines with AutoGluon / FLAML
Interpreting baseline results
Model Building
Linear Models
Linear regression for continuous numerical prediction
Logistic regression for classification
Generalized Linear Models (GLMs)
Ridge and Lasso regularization
Decision Trees
How decision trees split data
Key hyperparameters (max_depth, min_samples)
Preventing overfitting
Random Forests
Ensemble learning concept
Why forests outperform single trees
Extracting feature importance from Random Forests
Random feature subspace selection
Bagging (Bootstrap Aggregating) approach
Out-of-Bag (OOB) score and its use cases
When Random Forest excels
Model Building - Gradient Boosting Models
XGBoost
Understanding gradient boosting
Key hyperparameters
Regression and classification modes
Hands-on model building
Leveraging GPU acceleration in boosting
LightGBM
When to use LightGBM over XGBoost
Speed and memory advantages
Native categorical feature handling
CatBoost
Built-in categorical encoding
Ordered boosting to prevent overfitting
Comparing gradient boosting libraries
Evaluation Metrics and Loss Functions
Classification
Accuracy
Precision
Recall
F1-Score
ROC-AUC
PR-AUC
Regression
MAE
RMSE
MAPE
R²
Quantile loss
Inspecting Model Outputs
Learning curves for bias-variance diagnosis
Validation curves for hyperparameter sensitivity
Deep dive into the confusion matrix
Calibration plots and reliability diagrams
Error analysis techniques
Hyperparameter Optimization
Optimization Strategies
Grid search vs. Random search vs. Bayesian optimization
Hyperparameter tuning with Optuna
Multi-objective optimization
Pruning strategies for faster optimization
Which hyperparameters to tune first
Neural Networks (NN) for Tabular Data
Architectures and Use Cases
When to use neural networks for tabular data
PyTorch tabular models
State-of-the-art tabular architectures
Entity embeddings for categorical variables
Comparison with tree-based models
Imbalanced Datasets
Handling Class Imbalance
Class weight adjustments
SMOTE and its variants (BorderlineSMOTE, ADASYN)
Classification threshold optimization
TomekLinks
Cost-sensitive learning
Ensemble methods for imbalanced data
Model Explainability
Tools and Techniques
SHAP values: global and local explanations
LIME for local interpretability
Permutation importance
Partial Dependence Plots (PDP)
Individual Conditional Expectation (ICE) plots
Other Courses in This Category
-
Behavioural features, gradient boosting, anomaly scores and hybrid rule-model setups for payment, insurance and account fraud, with hands-on work on extreme class imbalance, delayed labels and operational threshold management.
- Live Course
- 5 modules
-
Variable analysis with WoE/IV, logistic regression scorecards, gradient boosting alternatives, reject inference and score calibration for credit, insurance and supplier risk, building explainable and auditable risk scores.
- Live Course
- 5 modules
-
Churn, purchase and cross-sell propensity models covering target definition, customer behaviour features, survival analysis, uplift modelling and campaign targeting, turning predictions into measurable retention and sales actions.
- Live Course
- 5 modules
-
Collaborative filtering, content-based and hybrid recommenders, matrix factorisation, implicit feedback, cold-start strategies and ranking metrics for building personalised product and content recommendation engines.
- Live Course
- 5 modules
-
Failure probability and remaining useful life (RUL) prediction from sensor, telemetry and maintenance logs, with hands-on signal feature engineering, health indices, early-warning thresholds and maintenance cost optimisation.
- Live Course
- 5 modules
-
SKU-location demand forecasting covering promotions, price elasticity, stockout censoring adjustment, intermittent demand (Croston/ADIDA), hierarchical reconciliation and probabilistic forecasts tied to safety stock.
- Live Course
- 5 modules
Get in touch
