Feature Selection in Imbalanced Data

The traditional feature selection methods are not suitable for imbalanced data as they tend to be biased towards the majority class. This problem is particularly acute in the field of medical diagnostics and fraud detection where the class distribution is highly skewed. In this paper, we propose a novel filter approach using decision tree-based F₁-score. The F₁-score incorporates the accuracy with respect to the minority class data and hence is a good measure in the case of imbalanced data. In the proposed implementation, the F₁-score is calculated based on a 1-dimensional decision tree classifier resulting in a fast and effective feature evaluation method. Numerical experiments confirm that the proposed method achieves robust dimensionality reduction and accuracy results. In addition, the low computational complexity of the algorithm makes it a practical choice for big data applications.