A hybrid model for handling the imbalanced multiclass classification problem

Abstrak

Data in many application domains is imbalanced. In machine learning, addressing imbalanced data is crucial to prevent bias towards the dominant class label and ensure that prediction models can learn and predict the minority class proficiently. This paper proposes a hybrid imbalanced classification model (HICD) to address the multiclass imbalanced data problem. The primary idea is to combine effective methods to construct a classification model that can handle multiclass imbalanced data effectively. Four methods are employed: an oversampling method to balance the data, a decomposition method to convert the multiclass problem into a set of binary problems, ensemble classification to integrate base classifiers to improve prediction, and a boosting method to encourage the classifier to pay more attention to misclassified samples. To evaluate the proposed model, seventeen imbalanced datasets from various application domains, featuring different numbers of classes, instances, features, and imbalance ratios, are assessed. The experimental results and statistical significance tests demonstrate that the proposed hybrid model significantly outperforms the standard one-vs-one (OVO) approach and the OVO combined with oversampling technique (SMOTE), both considered state-of-the-art for addressing imbalanced multiclass datasets, in terms of F1-score.