Handling highly imbalanced output class label: a case study on Fantasy Premier League (FPL) virtual player price changes prediction using machine learning / Muhammad Muhaimin Khamsan and Ruhaila Maskat

• Main Authors: Khamsan, Muhammad Muhaimin (Author), Maskat, Ruhaila (Author)
• Format: Book
• Published: Penerbit UiTM, 2019-12.
• Subjects: Article; PeerReviewed
• Online Access: Link Metadata
• Notes: https://ir.uitm.edu.my/id/eprint/61448/1/61448.pdf
• Keywords: Data processing

In practice, a balanced target class is rare. However, an imbalanced target class can be handled by resampling the original dataset, either by oversampling/upsampling or undersampling/downsampling. A popular upsampling technique is Synthetic Minority Over-sampling Technique (SMOTE). This technique increases the minority class by generating synthetic class labels and assigned the class based on the K-Nearest Neighbour (K-NN). SMOTE upsampling can only upsample at most one minority class at a time, which means for a multiclass dataset, it needs to undergo multilayer SMOTE to balance the class label distribution. This paper aims to find a suitable method in handling imbalanced class using dataset from Fantasy Premier League (FPL) virtual player to predict price changes. The cleaned dataset has a highly imbalanced class distribution, where the frequency of "Price Remain Unchanged (PRU)" is higher than "Price Fall (PF)" and "Price Rise (PR)". This paper compared between the baseline (original) dataset, SMOTE-applied dataset and shuffled, linear and stratified sampling in split train-test subset, based on a deep learning algorithm. This paper also proposed criteria of low values in standard deviation (distribution of true positive on each class label on accuracy) as a measurement for finding the best method in handling imbalanced class labels. As a result, multilayer SMOTE until all the classes distribution is the same, combined with stratified sampling in split training and testing subset, get the lower standard deviation (5.7873), high accuracy (80.06%) and less execution runtime (1 minute 41 seconds) compared to the original highly imbalanced dataset.