Imbalanced Learning (Unbalanced Data) - The Complete Guide

Apply data-based approaches to imbalanced learning by adjusting the class distribution through under sampling and oversampling, noting trade-offs like information loss and overfitting.

Explore undersampling methods that balance imbalanced data by removing majority-class instances, including random undersampling, near-miss, class-purity maximisation, and clustering-based approaches that preserve minority examples and address overlap.

Learn random undersampling, a simple technique that removes instances from the majority class to balance the dataset, while noting possible information loss about the majority class.

Explore random undersampling in imbalanced learning to balance data, reducing the majority class to 17 instances to match the minority, and visualize changes with scatter plots before and after.

Identify Tomek links as pairs of nearest neighbor instances from opposite classes and use them for under sampling to balance imbalanced data by removing majority points while keeping minority instances.

Demonstrate imbalanced learning with Tomek Link on the liver dataset, removing majority-class instances to balance the data, and visualize before and after with scatter plots.

Apply one site selection under sampling to balance data by removing unreliable majority instances while keeping all minority examples, using nearest neighbor and noise reduction.

apply a dataset example by loading data, inspecting size and features, separating X and y, and visualizing with scatter plots to compare the algorithm’s effect on majority and minority classes.

Apply class purity maximisation (CPM) under sampling by forming one minority and one majority center, partitioning data into two clusters via Euclidean distance and impurity comparison to the parent.

Cluster the data into K clusters, compute each cluster's majority-to-minority ratio, and remove majority samples to achieve a target ratio before combining with minority samples.

Delve into a practical clustering example by balancing an imbalanced dataset with a sampling algorithm, visualizing before and after results, and evaluating improvements with downstream models.

Learn how the edited nearest neighbor method cleans data by under sampling, removing majority-class instances whose labels differ from their k nearest neighbors while preserving minority samples in oversampling techniques.

Apply the edited nearest neighbor (ENN) method from the imbalance library to a real dataset, visualizing feature pairs and showing how ENN reduces the majority class to balance the data.

Balance imbalanced data using near-miss under sampling. Focus on near-miss 2 as the most widely used, with near-miss 1 targeting closest minority distances and near-miss 3 removing closest majority instances.

demonstrate near-miss version 2 on a liver dataset, balance the data from 345 to 290 samples, visualize with scatter plots, and compare results with and without sampling.

Increase minority instances through oversampling, not touching the majority. Explore methods like random oversampling, SMOTE variants (borderline SMOTE, safe level SMOTE), and adaptive synthetic sampling.

Increase the minority class by randomly sampling and adding minority class examples to balance the dataset, but this method can overfit by raising training accuracy without improving minority class generalization.

Apply a practical random oversampling example using the imbalanced library to duplicate minority samples and visualize before and after results on the dataset.

Learn SMOTE, a minority over-sampling method that generates synthetic samples along the line between minority instances using k-nearest neighbors, improving imbalance without replication.

Apply the SMOTE algorithm to balance an imbalanced dataset by generating synthetic minority samples, compare before and after visuals, and tune parameters like sampling strategy and neighbors.

Borderline SMOTE identifies dangerous borderline minority examples near the decision boundary and oversamples them by generating synthetic points along lines to nearest neighbors, strengthening the border region.

Apply the borderline-SMOTE algorithm to balance an imbalanced dataset, generating synthetic minority samples and comparing original versus augmented data with scatter visualizations.

Explore SMOTE-SL, a safe-region oversampling method that generates synthetic minority samples within detected safe areas using safe level and safe level ratio, aligning with nearest-neighbor distances.

learn how ADASYN adaptive synthetic sampling boosts performance on unbalanced data by adaptively generating minority samples according to their density distribution using k nearest neighbors, contrasting with SMOTE.

Apply the adaptive synthetic oversampling algorithm to an imbalanced data set, inspect dataset structure and class distribution, visualize feature pairs, and compare different majority–minority ratios to improve modeling.

Hybrid sampling blends minority oversampling with subsequent data cleaning to improve imbalanced data performance. Apply SMOTE or adaptive SMOTE, then nearest-neighbor or Tomek links cleaning for better results.

Apply SMOTE-ENN to an imbalanced dataset, oversampling minority and undersampling majority to balance classes. Visualize with scatter plots and compare different class ratios to assess improvements.

Demo of practical SMOTE-Tomek Link on an imbalanced dataset; explore the data, set X and y, visualize with scatter plots, apply the imbalanced library's hybrid oversampling, and compare results.

Explore the algorithmic level approach to imbalanced data, adjusting the existing model rather than resampling to help the minority class by learning from one class and tuning costs and parameters.

Explore cost sensitive learning by adjusting misclassification costs with cost matrices, guiding decision thresholds and learning processes for minority cases in imbalanced data, including decision trees and neural networks.

Explore cost sensitive learning on an imbalanced data set with a support vector machine, adjusting class weights to favor the minority class and interpret the decision boundary.

Explore one-class learning as a novelty detection method that builds a boundary around the target class using mainly its instances, addressing extreme imbalance and high dimensional noise.

Explore active learning as an iterative supervised approach for imbalanced data, selecting the most informative instances near the current hyperplane to retrain a kernel-based SVM with online learning.

Learn to evaluate classifiers on imbalanced data using standardized performance metrics and statistical tests. Build a confusion matrix, compute accuracy and recall, and interpret results to support conclusions.

Explore the confusion matrix for a two-class classifier, detailing true positives, true negatives, false positives (type 1 error), and false negatives (type 2 error).

Generate the confusion matrix and compute performance metrics from it on an imbalanced dataset, using an oversampling approach and a train/test split to evaluate a classifier.

Grasp accuracy and its limitations in imbalanced data, where minority class performance matters and misclassification costs differ. Reassess reliance on accuracy alone, which biases toward the majority class.

Learn to compute accuracy and error rate from a confusion matrix, and explain why accuracy can be misleading for imbalanced data.

Explore precision, recall, and specificity as key evaluation metrics for imbalanced learning, clarifying how precision measures exactness, recall (sensitivity) measures completeness, and specificity guards against false alarms.

Learn to compute precision and recall from true and predicted labels and the confusion matrix, and interpret sensitivity and specificity in imbalanced data contexts.

Explore F-measure, adjusted F-measure, and geometric mean for evaluating classifiers on imbalanced data, balancing precision and recall with sensitivity, specificity, and receiver operating characteristics.

Calculate the F1 score using a classifier's predictions and confusion matrix, and generate a classification report that shows recall, precision, F1, and micro, macro, and weighted averages.

Import the geometric means core function and apply it to the class and protected class to compute the geometric mean score. The example reports a score of 0.85 (85%).

Use ROC and AUC to assess imbalanced data performance, visualizing true positive rate against false positive rate across thresholds. Include one-vs-all extensions for multi-class.

Calculate the ROC AUC score using class labels to compute 0.859, and learn how this metric informs final decisions and recommendations.

Compare unbalanced dataset performance using nonparametric methods, including Iman-Davenport and Wilcoxon tests, Friedmann ranking on AUC results, and post hoc analyses for algorithm comparisons.

Explore the difference between underfitting and overfitting, why models may fail on new data, and how train-test split helps detect and prevent overfitting in machine learning.

Split data into training and testing sets for imbalanced data, using a random split to ensure class representation and evaluate future performance.

Learn to use a validation set to tune your model with training data, evaluate on validation only, and then test on the test set to avoid overfitting.

Explains cross validation as an extension of the train-test split, using k folds to train on k-1 subsets and test on the remaining one, with bias-variance tradeoffs and fold recommendations.

Explore practical, step-by-step strategies to handle imbalanced data, from data collection and simple resampling to synthetic sampling and hybrid methods, with performance metrics, evaluation, and reproducibility at the core.