Tuning Endpoint-variability Parameters by Observed Error Rates to Obtain Better Prediction Accuracy of Pointing Misses

Abstract

Error rates (ERs) in target-pointing tasks are typically modelled in two steps: predicting the click-point variability (σ) based on target sizes and then computing the probability that a click falls outside a target. This is an indirect approach if the researcher's purpose is to achieve the accurate prediction of ERs because the model coefficients are optimized to predict σ accurately in the first step. We compared the prediction accuracies of this method with a more direct technique in which the coefficients used for σ are determined in such a way as to optimize the closeness between observed and predicted ERs. Our re-analysis of eight datasets from mouse- and touch-based pointing studies showed that the latter approach consistently outperforms the conventional one if the starting values for the parameter search are appropriate (which can be achieved by hyperparameter optimization), thus enabling the interface configuration on the basis of accurately predicted ERs.