Hybrid Sampling Approach for Multiclass Imbalanced Data

Machine learning algorithms may create skewed results in classification. In some of the scenarios, like machine fault detection, fraud detection, and disease diagnosis, the intent class or the focused class instances may be very few when compared to the other instances or non-focused instances. In these cases, the classifier is trained based on the other non-focused classes and creates a skewed result. This will be high for false negatives. This paper proposes Stratified Near Miss Undersampling with Deep Neural Network (DeepSNMU) to investigate the imbalanced dataset using the Spark framework. The proposed DeepSNMU method uses IQR outlier detection for pre-processing, near-miss undersampling for balancing the dataset, and a Deep Neural Network for classification. Experiments were conducted using four highly imbalanced datasets from the KEEL repository. These datasets have multiple target classes, which are experimented with using DeepSNMU. In this research work, DeepSNMU produces high classification accuracy. The result of DeepSNMU is compared with existing balancing techniques and classifiers. The outcome of this research work has shown higher accuracy in predicting the minority intent class than the existing methods.