This study proposes a novel data analysis framework that utilizes pattern grouping to identify similar and contrasting subset gesture groups. We start by forming an initial group of gestures, subsequently expanding it incrementally based on pairwise distances to the group’s centroid. The evaluation of pairwise distances allows the group to remain coherent and consist of highly similar gestures. The process continues until the classifier model’s accuracy reaches its lowest point, indicating the introduction of dissimilar gestures. We then assign these selected gestures to a similar group and the remaining gestures to the contrasting group. Afterward, we train separate classifiers for each subset group. The separated training allows the model to learn the similar gestures so that it can have higher separation ability. The proposed framework was tested under five classifiers: artificial neural networks (ANN), k-nearest neighbors (kNN), random forests (RF), support vector machines (SVM), and XGBoost. We fine-tuned the classifiers to ensure the optimality of each classifier. Each classifier was run 10 independent times, and the results were aggregated for final accuracy reporting. We found that the proposed framework increases the accuracy level of the five tested classifiers significantly. Notable improvements were noticeable in the ANN and SVM, with improvements from 77.98% to 87.26% and from 60.03% to 67.74%, respectively. It also means that the proposed framework allows more extensive learning to recognize the discrepancy of the gestures.
