Hang Zhao will be presenting his PhD thesis proposal on fundamental interactions on Wednesday.
Title: Probabilistic Modeling of Fundamental Interactions and Their Applications
Zoom Link: https://stonybrook.zoom.us/j/93646975829?pwd=r9oC2TIVvGiSOBOVnkcTDDoTgeVLWf.1
Zoom ID: 93646975829
Passcode: 992911
Abstract
Basic interaction models such as Fitts' law, Steering law, and Finger-Fitts law are cornerstones for interface development, optimization, and evaluation. Despite the prominence, however, these basic interaction models are deterministic, which predict only the mean movement time (MT) of interactions, but provide no information on the variability or the probability distribution of MT. This dissertation addresses the challenge by creating variance, distribution, and Bayesian hierarchical models for basic interactions including pointing, steering, and applying these models to solve a real-world problem: estimating the probability of an older adult developing Parkinson's Disease (PD) by analyzing the user's interaction with computers and mobile devices.
Firstly, we introduce the quadratic variance and distribution models that predict variances and distributions of MT in trajectory-based steering tasks. The quadratic variance model shows that the index of difficulty of a task is quadratically related to MT and distribution models are built using the Steering law to predict the mean, the quadratic model to predict the variance, and distribution models such as Exponential Modified Gaussian (exGaussian) to predict the distribution of MT, enhancing predictions from point estimates to variance and distribution estimates.
Secondly, we extend Fitts' law into a Bayesian hierarchical model to predict MT distributions in cursor-based target acquisition. We adopted the exGaussian distribution with Fitts' law predicting the mean and the quadratic variance model predicting the variance from our first study under the 2-stage hierarchical architecture to build the Bayesian hierarchical models.
Thirdly, we develop a mobile game named MoleBuster to gather data from users with and without PD. Using this game, we construct temporal and spatial models based on touch-pointing data from these users. This information, along with our newly designed CNN-Transformer models, aids in detecting the presence of PD.
Lastly, we utilize data from mouse-based pointing and steering actions, obtained from our prior studies on computer devices, and build Bayesian hierarchical models as well as SVM, LSTM-CNN, CNN-Transformer, and the combination models to detect PD.
