Machine Learning Essentials (Lecture Notes)

Vocabulary flashcards covering ML applications, learning types, data splitting, cross-validation, and Scikit-Learn pipelines.

Applications of Machine Learning

Common applications include healthcare (disease prediction), finance (fraud detection), e-commerce (recommendations), autonomous vehicles, NLP (chatbots, translation), and computer vision (facial recognition).

Supervised Learning

Training with labeled data to predict outputs. Examples: Linear Regression, SVM, Decision Trees.

Unsupervised Learning

Training with unlabeled data to find hidden patterns. Examples: K-Means, PCA, Clustering.

Reinforcement Learning

Learning through trial and error with rewards/punishments. Examples: Q-Learning, AlphaGo.

Batch Learning

Learns from the entire dataset at once; retraining needed for updates.

Online Learning

Learns incrementally from data streams; adapts continuously.

Overfitting

Model fits training data too well but fails on new data.

Regularization

Technique to prevent overfitting by penalizing model complexity.

Underfitting

Model is too simple, fails to capture patterns.

Training Set

Data used to train the model.

Testing Set

Data used to evaluate model performance on unseen data.

Dataset Split (70–80% / 20–30%)

Typical split: 70–80% training data and 20–30% testing data.

K-Fold Cross Validation

Data is split into k folds; train on k−1 folds, test on the remaining fold; repeat k times.

Stratified Sampling

Ensures class proportions are preserved in train/test splits (important for imbalanced datasets).

Scikit-Learn Design

Main features: Consistent API, Estimators (fit, predict, transform), Transformers, Pipelines, Cross-validation tools, Metrics.

Estimator (in Scikit-Learn)

An object with methods like fit, predict (and transform) used to fit models.

Transformer

An object that transforms data (used inside Pipelines).

Pipeline (Scikit-Learn)

A sequence of preprocessing + model steps applied consistently.

Why Use Pipelines

Prevents data leakage, simplifies workflows, and ensures transformations apply to both training and testing.

Pipeline Example

StandardScaler → Logistic Regression model (a typical pipeline).

Generalization error

The error rate on new cases

Overfitting the training data

If the training error is low but the generalization error is high, it means that your model is…

Fit

The ______method is used to build models

Association Rule Learning

Discover patterns and relationships between attributes in large datasets

Semisupervised Learning

Few labeled instances and plenty of unlabeled instances, Combinations of unsupervised and supervised algorithms