Enhancing the performance of kNN for glass identification dataset using inverse distance weight, ReliefF ranking and SMOTE.

Most real-world classifications problem provides an uneven, skewed dataset. Imbalanced datasets are a key issue in machine learning. The challenge for any machine learning algorithm to learn from imbalanced dataset is correctly classifying cases from the minority class. Imbalanced dataset leads machine learning algorithms to misclassify cases from the minority class which is undesirable for real world classification applications and also degrades the performance of algorithm. Glass identification dataset is one of the datasets that serve as a example of skewed imbalance real world dataset. The majority of research on glass identificationtasks ignores the data outlier and class imbalance found in this dataset. Class imbalance, data outlier, and insignificant featuresare only a few of the factors that contribute to the poor performance of distance-based machine learning methods such as kNN. Thus, this paper explores the performance of one of the most simple and efficient machine learning classifier kNN. The authors performed interquartile range (IQR) thresholding to eliminate data outliers, ReliefF ranking to find significant features and SMOTE to cope with the class imbalance occurring in the dataset. The Inverse distance weighted kNN is utilized for classification in the improved dataset. The results revealed that using the strategies mentioned improves the performance of the kNN algorithm while also providing accuracy that is better and more dependable than much of the research done in this dataset. In this paper the authorsfound the highest accuracy of glass identification task is 78.923% with F-measure 0.791. [ABSTRACT FROM AUTHOR]