The goal of this project is to develop the principles to design a virtual personality assessment laboratory. This requires development of a preliminary taxonomy of mechanics, grounded on personality research, that allows researchers to use behavioral patterns of individuals in computer-generated virtual environments, to assess their real-world personality. Such virtual personality detection mechanisms can then be used by other researchers to adapt the laboratory further, which would be one ultimate goal: the design of more personalized and adaptive applications that may improve impact on large societal problems. The research activity will lead to a new methodology with the potential to transform current practices across disciplines, from social psychology to human-centered computing.
The virtual personality assessment laboratory will be developed as a set of modular challenges and situations that make use of the mechanics individuated in the taxonomy. These situations are constructed to elicit personality preferences. The design is driven by personality theory and validated by a wide range of personality measures such as the Need For Cognition, the California Q-Sort, the Reiss Motivation Profiler and the Five Factor Model. The system will be validated through two iterations to ensure that scenarios are assimilated and that they conform to the intention of the designers. A final summative evaluation will be administered utilizing data on the behavior of research subjects inside the environment, as well as various personality measures such as scores from personality questionnaires, informant interviews, and behavior coding. Correlation analysis will be used to investigate relationships between choices emerging from the context of action in the virtual environment and personality scores.
This research impacts directly a number of disciplines from psychology of personality to adaptive technologies and personalization. The research affects our understanding of personality within virtual environments, which are becoming a major part of our lives. It also has the potential of developing customizable learning environments. Understanding individual differences through the analysis of consumption and behavior of digital entertainments allows for a deeper level of adaptation and personalization of persuasive technologies aimed at fostering education, health or training, potentially increasing participation from all segments of society.
Our principal goal for this research project was to explore how an individual’s personality impacts their behavior within virtual worlds. This topic is of potential interest to virtual world designers and developers, as it provides one potential method to broaden the appeal of their product to a wide range of customers. However, the connection between behavior in virtual environments and personality is also very interesting to psychologists. Current methods to assess personality have either limited reliability (e.g., internet surveys) or require a great deal of time and effort (bringing a person into the lab for a 2+ hour evaluation), and a specialized virtual environment that could aid the process would be a valuable tool. We first designed a custom scenario using the Bethesda Softworks game, Fallout: New Vegas (Figure 1). The scenario is experienced from the first-person perspective, and is primarily a role-playing experience. Users can explore the world, pick up items and equipment, talk to computer-controlled characters (NPCs), perform quests and occasionally engage in combat. Their movements and interactions are automatically logged, which forms the behavioral data set. For the experiment they were also required to take an online personality test (NEO-PI-R) which scored their personality across five primary traits and 30 secondary facets (six per trait). Figure 1. Screenshot from the game where the user in in conversation with an NPC and must select a response. There were several areas in the game, each containing different behavioral affordances (opportunities for action). Figure 2 shows an overview of how users’ behaviors changed according to the area they were in. We found several reliable correlations between user behavior and personality when we broke down the data by area. For example, the amount of user movement in the outdoor region was correlated with the trait extraversion, and the amount of time spent talking to NPCs in the hotel areas was correlated with the facets achievement-striving and self-discipline. Figure 2. Summary of behaviors in the Virtual Environment for all subjects, broken down by location. In some cases, the correlations were reliable enough to work backwards and estimate a user's personality score from their behavior. Figure 3 summarizes prediction errors of a model for predicting the agreeableness trait based on how many times the user talked to a virtual agent, how much time they talked and how much time they spent performing quests, all in the starting area of the game ("Intro House"). The model was able to predict the agreeableness of 76% of the subjects to within 15% accuracy, or 88% of subjects to within 25%. Work is ongoing to find improved behavioral measures that increase predictive power. Figure 3. Breakdown of prediction errors for estimating Agreeableness from game behavior. In a subsequent study, we further divided the game areas into microlocations (Figure 4) to define the affordances more precisely. For example, location 01 contained only bathroom fixtures (sink, toilet, and bathtub), location 02 had the NPC "Tracy" and location 04 contained a dresser with several objects to pick up. For the behavioral measure, we used the amount of time spent in a microlocation. This can be visualized using a heatmap (Figure 5) that shows the cumulative time all users spent in each part of the Intro House area. Figure 4. The layout of the Intro House game area with microlocations indicated. Figure 5. Heatmap of all users' position data in the Intro House. The second study also treated personality holistically. In other words, rather than looking at single traits for the users, we looked at all five traits simultaneously. Userswere grouped using a technique called hierarchical clustering, which defines a set of "typical" personality types and assigns users to those types based on similarity. After classifying the users, we looked at whether they spent more or less time than the average in the microlocations defined earlier. As Table 1 shows, each archetype has a characteristic set of locations that its members tend to prefer (++, +) or avoid (--, -). Table 1. Time spent in Intro House locations (columns) according to personality archetype (rows). The plusses indicate that archetype members tended to spend more time in a location; minuses indicate they spent less time there. Locations 6 and 7 were not preferred or avoided by any archetype. To summarize, an individual’s personality appears to make a deep imprint on their behavior not just in real life, but in the digital realm as well. The environment or context in which they find themselves is crucial as well – different behaviors are warranted for a spacious outdoors versus a cramped indoors, or a friendly versus hostile environment. We are continuing to develop techniques to extract behaviors from the game logs, and account for other factors such as the state of the game or what the user has previously accomplished.
