Applied Machine Learning with Python

Translate a business question into a learning problem and establish a baseline before modeling.

• Supervised versus unsupervised learning
• Classification, regression, and clustering use cases
• Defining the target and the prediction moment
• Train, validation, and test splits
• Baseline models such as majority class and mean prediction
• The scikit-learn estimator API: fit, predict, transform

Hands-on lab: Take a business brief, define the target and success measure, and build a dummy-model baseline that later models must beat.

Train core model families and understand when each is a sensible choice.

• Linear and regularized regression with Ridge and Lasso
• Logistic regression and probability outputs
• Decision trees and their instability
• Random forests and bagging
• Gradient boosting with XGBoost and LightGBM
• k-nearest neighbors and the effect of feature scaling

Hands-on lab: Train five model families on one dataset, compare them against the baseline, and write a short note on which you would ship and why.

Prepare features reliably and package preprocessing with the model.

• Imputing missing values
• One-hot, ordinal, and target encoding
• Scaling and transforming skewed features
• Date, text-length, and interaction features
• ColumnTransformer and Pipeline
• How preprocessing outside a pipeline causes data leakage

Hands-on lab: Rebuild a notebook with leaky preprocessing as a single Pipeline, and measure how much the reported score changes.

Evaluate models honestly, recognize when performance will not hold up in production, and tune without fooling yourself.

• Bias, variance, and learning curves
• k-fold, stratified, grouped, and time-series cross-validation
• Classification and regression metrics such as F1, PR AUC, MAE, and RMSE
• Handling class imbalance and choosing thresholds
• GridSearchCV, RandomizedSearchCV, and Optuna search
• Nested cross-validation and tracking runs with MLflow

Hands-on lab: Evaluate an imbalanced classifier with the wrong and right metrics, select a threshold with a written justification, then tune a gradient boosting model with Optuna and log every run in MLflow.

Find structure in unlabeled data using clustering, dimensionality reduction, and anomaly detection.

• k-means and choosing k
• Hierarchical clustering and DBSCAN
• Evaluating clusters with silhouette score and business sense
• Principal component analysis
• t-SNE and UMAP for visualization only
• Anomaly detection with Isolation Forest

Hands-on lab: Segment a customer dataset with two clustering methods, profile the segments in plain language, and flag anomalies in a sensor dataset.

Explain model behavior globally and for individual predictions, and check models for fairness before they affect people.

• Permutation importance versus built-in importance
• Shapley values and the intuition behind SHAP
• Global summary and dependence plots
• Local explanations with waterfall plots
• Sources of bias, proxy variables, and group-level fairness metrics
• Model cards and documentation

Hands-on lab: Explain a boosted-tree model with SHAP and write a plain-language explanation of one prediction, then audit error rates across subgroups with Fairlearn and draft a model card stating limits and intended use.

Prepare a model so someone else can reproduce, run, and monitor it.

• Saving pipelines with joblib
• Reproducibility: seeds, versions, and environment files
• A simple prediction function or API with FastAPI
• Input validation and schema checks
• Monitoring for data drift and performance decay

Hands-on lab: Save your final pipeline, serve it behind a small FastAPI endpoint, and write a README covering inputs, outputs, and monitoring suggestions.