An AI Model to Predict Dominance, Nervousness, Trust, and Deception from Behavioral Features in Videos
Author
Walls, Bradley L.Issue Date
2020Keywords
Artificial IntelligenceBehavioral Features
Deception Detection
Emotion Recognition
Predictive Modeling
Advisor
Nunamaker, Jay F.Burgoon, Judee K.
Metadata
Show full item recordPublisher
The University of Arizona.Rights
Copyright © is held by the author. Digital access to this material is made possible by the University Libraries, University of Arizona. Further transmission, reproduction, presentation (such as public display or performance) of protected items is prohibited except with permission of the author.Embargo
Release after 12/21/2022Abstract
When leveraged strategically, the perception of the degree to which an individual demonstrates a relational message theme plays a vital role in determining the success of key outcomes within specific contexts. For example, in the fields of leadership, negotiations, and conflict resolution, knowledge of the revealing patterns of dominance, as well as knowing how to project dominance throughout on-going group interactions, increases the probability that a person engaging in deliberate actions will achieve desired results. For this research, an interactive group game was designed and played with groups of six to eight participants. In each game, videos of all participants were recorded and analyzed as players employed various behaviors (i.e., dominance, trust, deception, etc.) to be elected as mission leaders and manipulate others to win the game. Analysis of participant videos required the extraction of face landmark locations, facial Action Units (AUs), and facial rigidity assessments. These extracted features are turned into three hundred seventy micro-behaviors generated for every frame of video of 182 players over 26 games. Dimensionality reduction is accomplished using neighborhood component analysis (NCA), and a supervised classification model is trained using in-game survey data from participants.Analysis of balanced experimental data shows using an ensemble classification method for 2-way categorical classification of six perceptual judgments attains the following performance levels: • Dominance prediction achieves an accuracy of 77.55% • Not-dominant prediction achieves an accuracy of 84.52% • Nervousness prediction achieves an accuracy of 84.66% • Not-nervous prediction achieves an accuracy of 71.93% • Trust prediction achieves an accuracy of 77.10% • Not-trust prediction achieves an accuracy of 79.56% In each case, a unique set of top 20 best discriminating micro-behaviors are used to directly predict the corresponding perceptual judgment. In the case of attributions, classification is evaluated from both a direct and indirect approach. Using a Brunswik Lens model as a guide, 2-way categorical classification accomplished these top performance levels: • Deception prediction achieves an accuracy of 71.05% • Leader prediction achieves an accuracy of 71.44% • Winner prediction achieves an accuracy of 65.84% Detailed inspection of top-performing features does not reveal any single feature as an absolute indication for discrimination. It is the subtle combination of a set of features that provides predictive performance. This suggests that technology-assisted solutions should be the preferred approach over exclusive human training to accomplish these types of assessments. "The beauty comes not from a single flower, but from the garden as a whole."- Kellie KeelType
textElectronic Dissertation
Degree Name
Ph.D.Degree Level
doctoralDegree Program
Graduate CollegeManagement Information Systems