Instructions for Preparing Papers for Proxies:Jurnal Informatika

Abstract

The aim of this study is to introduce a hybrid machine learning solution to classify athletes' fitness levels using unlabelled physical activities using clustering and classification techniques. This work focuses on overcoming the challenge posed by datasets of fitness levels which lack defined fitness levels labels required for direct classification processes. The selected dataset in this paper is the Workout & Fitness Tracker Dataset which consists of about 10,000 samples with selected features such as Steps Taken, Calories Burned, Distance, Workout Duration, and Daily Calories Intake. In the first step, K-Means and K-Medoids clustering techniques were implemented to form clusters in a certain number of cluster formations (K=3, K=4, K=5, K=6), then generating the pseudo-labels. Clustering effectiveness was analysed by evaluating the Silhouette Scores. Finally, in the second step, the generated pseudo-labels were fed into the Modified K-Nearest Neighbour (MKNN) classification algorithm. The experiments have demonstrated that the K-Means clustering technique provided the maximum Silhouette Score of 0.1476 with the formation of K=6. On the other hand, K-Medoids clustering provided K=3 and resulted in 94.80% accuracy and 94.79% F1-Score. It should be mentioned that clustering efficiency may not guarantee the highest classification accuracy.