20.08.2024

Models to predict human and machine behavior

The University of Tübingen Methods Center

“There’s a great need to model human behavior using what is known as soft data, so that it can more reliably be studied empirically,” explains Professor Augustin Kelava, director of the University of Tübingen’s Methods Center. “Describing human behavior is not the problem. But developing models that enable the prediction of future behavior is a deficit that we are working to confront with the Methods Center at Tübingen,” says Kelava. With a doctorate in psychology, Augustin Kelava is the Professor of Quantitative Methods and has been the director of the Methods Center since its foundation in 2018. The Methods Center is both an institutional part of the Faculty of Economics and Social Sciences and a core facility of the University of Tübingen, which is co-funded from the Excellence Strategy. A crucial aspect in its foundation was the aim to improve links between the social sciences and natural science, technical and statistical disciplines. 

Quantitative Data Science degree program

Connecting the three subjects - psychometry, econometry and machine learning - gave rise to the Quantitative Data Science degree program in 2020, which is now on its fourth intake of students. “They’re fantastic graduates, who will move on for example to careers at the Max Planck Institute or to industry such as Bosch or Mercedes-Benz. Here at the Methods Center we’ve been able to fill a gap both in teaching and in research, and are continuing to expand possible cooperations,” reports Kelava.

“Our unique feature in research is that we develop, for example, procedures for the forecasting of complex time series, which are transferable to people and otherwise tend to be developed and used in signal processing. We’re now able to compete in research and teaching with international locations in the USA,” says Kelava. As a service for scientists, the Methods Center team offers advice and training. Requests come for example from colleagues in sociology or psychology for advice on statistics or from staff at the Max Planck Institute for psychometric measuring tools. Proactively, the Methods Center offers workshops, such as Spring Schools or Fall Schools, which offer methods training over several days or a week.

Predicting dropouts 

One of the Methods Center’s research projects studies the modeling of student dropouts. “We’re not interested in the rates of dropout, as that is fundamentally the end result of a process. That’s what has been done as standard worldwide for some time now: identify rates of dropout and then look for causes. By contrast, what we’re doing is developing statistical models that allow an individual prognosis,” explains Augustin Kelava. 

Models like this are also used in econometrics or engineering. The approach is that there are ‘intra-individual’ differences between people, e.g. dispositions. These can include for example how good someone was at mathematics in school or what their level of abilities was on admission. “We combine these data with variables that change over time. We’re already looking in routine surveys during the study introduction phase in the first semester at how someone feels, or obtaining self-assessments, which vary. From both these sources of information, we draw conclusions in real time about something that’s not yet apparent: the intention to give up studying,” says Kelava.

With this data, it is possible to identify the risk when the students are still attending university and taking part in their courses. This allows the teachers to respond quickly and work to avoid them dropping out. A University of Tübingen Methods Center team is currently programming software which they hope to make widely available as a service, so that universities can adapt and use it. 

The ‘personality’ of Large Language Models

“We’re cooperating with the Max Planck Institute for Intelligent Systems on a research project which looks at the personality of Large Language Models,” reports Augustin Kelava. This concept of ‘personality’ is often discussed in psychology, however there are different measurement methods for this, i.e. options for recording the experience of habitual human behavior via questionnaires. 

“Researchers at large international companies such as Google are applying these processes to machines and trying to record how extrovert, agreeable, etc., Large Language Models are. There are scientific publications on this already,” explains Kelava. However, the same measuring tools are applied to the machines as to humans. Kelava and his team show in their publication that this procedure cannot simply be transferred to machines. “That means we can’t draw any conclusions from the answers that the machines provide in the questionnaires. In principle, the way machines place their cross in a personality questionnaire differs completely from how humans do it.”

Since people are increasingly interacting with machines, Kelava regards the ‘psychology of machines’ as a major subject for research in future. “How do machines function psychologically and how can I record this? Questions like this are what we at the Methods Center really want to research.”

Johannes Baral