1. A larger dataset might be more balanced on the classes

2. More examples of minor classes may be useful in resampling

Accuracy may mislead in an imbalanced dataset

Metrics: confusion matrix, precision, recall, F-1, Kappa, ROC curves, etc.

1. Over-sampling: boostrap resample with replacement instances of under-represented class

2. Under-sampling: delete instances from the over-represented class

Randomly sample the attributes from instances in the minority class

1. Generative Models, Naive Bayes, KDE

2. Synthetic Minority Over-sampling Technique (SMOTE)

• Selects two or more similar instances, and perturb an instance attribute at a time by a random amount within the difference to the neighboring instances

Tree models often perform well on imbalanced dataset

Imposes a penal cost on the model for making classification mistakes on the minority class during training

1. Anomaly detection, take the instances in the minor class as the outliers

2. Change detection, detecting anomalous sequences/states in a time series

Decompose your larger class into smaller number of other classes

Classification when 80% of my training set is of one class at Reddit

A Brief Introduction to Change Point Detection using Python

Cohen's kappa

8 Tactics to Combat Imbalanced Classes in Your Machine Learning Dataset