Machine Learning with Python

What machine learning is and the vocabulary used throughout the field.

• Learning from data versus hand-written rules
• Supervised, unsupervised and reinforcement learning
• Features, labels, training and test data
• Regression versus classification
• The typical machine learning workflow
• Overfitting and underfitting at a glance
• Setting up Python, NumPy, pandas and scikit-learn

Covered in 1:1 mentoring sessions

The maths and preparation steps that every later module relies on.

• Vectors, matrices and dot products
• Derivatives and the gradient
• Probability and distributions refresher
• Handling missing values
• Encoding categorical features
• Feature scaling
• Train, validation and test splits
• Data leakage

Covered in 1:1 mentoring sessions

The simplest and most useful model, and the home of gradient descent.

• Simple and multiple linear regression
• The cost function: mean squared error
• Gradient descent and learning rate
• The normal equation
• Polynomial features
• Assumptions and residual plots
• R-squared and adjusted R-squared

Covered in 1:1 mentoring sessions

Keeping models simple enough to work on new data.

• Bias and variance
• Overfitting and underfitting
• Ridge (L2) regression
• Lasso (L1) regression and feature selection
• Elastic net
• Learning curves
• Choosing the regularisation strength

Covered in 1:1 mentoring sessions

Predicting categories and probabilities.

• The sigmoid function and odds
• Log loss and training
• Decision boundaries
• Multiclass classification: one-vs-rest and softmax
• Thresholds and class probabilities
• Class imbalance and class weights
• Confusion matrix

Covered in 1:1 mentoring sessions

Two simple classifiers built on distance and probability.

• Distance measures
• Choosing k
• The curse of dimensionality
• Why scaling matters for kNN
• Bayes' theorem for classification
• Gaussian and multinomial naive Bayes
• Text classification with bag of words

Covered in 1:1 mentoring sessions

Models that learn a series of questions from data.

• Splitting on features
• Gini impurity and entropy
• Information gain
• Controlling depth and pruning
• Regression trees
• Feature importance
• Strengths and weaknesses of trees

Covered in 1:1 mentoring sessions

Combining many models for better predictions.

• Why ensembles work
• Bagging and bootstrap samples
• Random forests
• Out-of-bag error
• Boosting and AdaBoost
• Gradient boosting
• Stacking and voting
• Comparing forests and boosting

Covered in 1:1 mentoring sessions

Maximum-margin classifiers and the kernel idea.

• Margins and support vectors
• Hard and soft margins
• The C parameter
• The kernel trick
• Linear, polynomial and RBF kernels
• Gamma and tuning
• SVM for regression
• When to choose an SVM

Covered in 1:1 mentoring sessions

Finding groups in data without labels.

• Unsupervised learning goals
• k-means and choosing k
• The elbow method and silhouette score
• Hierarchical clustering and dendrograms
• DBSCAN and density-based clustering
• Gaussian mixture models overview
• Customer segmentation example

Covered in 1:1 mentoring sessions

Compressing many features into a few useful ones.

• Why reduce dimensions
• Principal component analysis step by step
• Eigenvectors and explained variance
• Choosing the number of components
• PCA for visualisation and speed-up
• t-SNE overview
• Feature selection versus extraction

Covered in 1:1 mentoring sessions

Measuring models honestly and choosing between them.

• Cross-validation: k-fold and stratified
• Regression metrics
• Precision, recall, F1 and accuracy
• ROC curve and AUC
• Precision-recall curve for imbalanced data
• Grid search and random search
• Baselines and statistical caution
• Model interpretation basics

Covered in 1:1 mentoring sessions

Putting preprocessing and models together and finishing a real project.

• Pipeline and ColumnTransformer
• Preventing leakage with pipelines
• Tuning a whole pipeline
• Saving models with joblib
• Simple model deployment idea
• Project: from raw data to final model
• Reporting results and limits

Covered in 1:1 mentoring sessions