Toggle navigation sidebar

Toggle in-page Table of Contents

Titanic Survival Machine Learning

Titanic Survival

Setting up an environment file to run these notebooks

Setting up an environment

Data preprocessing

Kaggle Titanic survival - data preprocessing

Basic methods

Logistic regression model
Measuring model accuracy with K-fold stratification
Avoiding over-fitting with regularisation
Accuracy measurements in machine learning
Application of alternative accuracy measurements to a logistic regression model
Learning curves - how much data do we need?
Receiver Operator Characteristic (ROC) curve
Checking model calibration
Measuring uncertainty in probability predictions using bagging

Optimising machine learning models

Optimising machine learning models with Optuna

Feature selection and expansion

Feature selection using univariate statistical selection
Feature selection using forward selection
Feature selection using backward elimination
Feature expansion

Working with imbalanced data sets

Dealing with imbalanced data by model weighting
Dealing with imbalanced data by under or over sampling
Dealing with imbalanced data by changing classification cut-off levels
Dealing with imbalanced data by enhancing the minority class with synthetic data (SMOTE: Synthetic Minority Over-sampling Technique)
The effect of over-sampling and under-sampling on model calibration

Random forest and XGBoost models

A Random Forest model
Random Forest Receiver Operator Characteristic (ROC) curve and balancing of model classification
XGBoost as a replacement for Random Forest

TensorFlow neural nets

TensorFlow neural net
TensorFlow api-based neural net.
TensorFlow Receiver Operator Characteristic (ROC) curve and balancing of model classification
TensorFlow ‘Wide and Deep’ neural nets
TensorFlow Bagging

Pytorch neural nets

PyTorch simple sequential neural net
PyTorch class-based neural net
PyTorch class-based neural net using GPU if avaliable

Model explainability

Explaining model predictions with Shapley values - Logistic Regression
Explaining model predictions with Shapley values - Random Forest
Explaining model predictions with Shapley values - Neural Network
Examining interactions between features with SHAP interactions

Other

Creating synthetic Titanic passenger data with SMOTE
A comparison of calibration of neural networks using a single sigmoid output or dual SoftMax or Sigmoid outputs

Powered by Jupyter Book

Index

By Michael Allen & Kerry Pearn
© Copyright 2022.