A Study On Force Pattern-Based Input Data Classification Using Machine Learning

This paper presents a personalized force-pattern input interface that enables multiple commands from a single touch location. Instead of mapping force to discrete levels or a continuous value, we capture the full force trajectory during a touch and treat it as a time-series “force pattern.” From each pattern we compute 13 statistical features (e.g., mean, standard deviation) and train machine-learning classifiers. Data were collected with an FSR sensor: four participants performed four self-defined patterns repeatedly, producing 640 samples. In a train–test split over the full dataset, the method achieved 0.9654 average accuracy. To account for distribution changes across collection dates, we additionally performed five-fold cross-validation within each date; Random Forest showed the most stable performance. We also found that feature scaling materially affects accuracy: Standard, MinMax, MaxAbs, and Robust scaling generally improved results, whereas vector normalization performed worse. When reducing the number of training samples per pattern, accuracy decreased overall, but some models remained robust; notably, Logistic Regression and SVM exceeded 0.8 accuracy even in low-data settings. Finally, under a broader condition that jointly classifies user and action, certain model–scaler combinations reached roughly 0.9 accuracy, suggesting applicability beyond strictly personalized use. These findings support practical multi-command touch interaction with modest calibration.