Trust in Human-AI Interactions
Artificial intelligence (AI)-based systems are altering the way humans, companies, and governments work, interact, or conduct business. Despite the benefits of introducing AI into the fabrics of society, their adoption faces various challenges and an increased need for human-centered design approaches that foster trust.
This project aims at exploring
– How can we define trust in digital interactions? And how to measure it?
– What fosters trust in digital interactions? And what does not?
– How can explainable AI and transparency improve human-AI interactions? And how can they not?
The project addresses these questions in the context of different digital interactions, including with data scientists, costumers, and domain experts.
A better understanding of how to foster trust in various digital interactions and how the design of digital interactions can be improved.
Selected Scientific Contributions:
Benk, M., Kerstan, S., Ferrario, A. (2023). You haven’t changed a bit! Initial Findings from a Bibliometric Analysis of Two Decades of Empirical Trust in AI Research. ACM CHI 2023 Workshop on Trust and Reliance in AI-Assisted Tasks (TRAIT’23)
Benk, M., Weibel, R. P., Feuerriegel S., & Ferrario, A. (2022). “Is It My Turn?”: Assessing Teamwork and Taskwork in Collaborative Immersive Analytics. Proc. ACM Hum.-Comput. Interact. 6, CSCW2, Article 479 (November 2022)
Benk, M., Weibel, R. P., & Ferrario, A. (2022). Creative Uses of AI Systems and their Explanations: A Case Study from Insurance. ACM CHI 2022 Workshop on Human-Centered Explainable AI (HCXAI’22).
Benk, M., Tolmeijer, S., von Wangenheim, F., & Ferrario, A. (2022). The Value of Measuring Trust in AI-A Socio-Technical System Perspective. ACM CHI 2022 Workshop on Trust and Reliance in AI-Human Teams (TRAIT’22).
Michaela Benk, Joseph Ollier
Young drivers are overrepresented in car accidents. One cause is their overestimation of their driving abilities due to a lack of feedback. New telematic applications provide a unique opportunity for giving drivers feedback. SaferDrive aims at making feedback more attractive for different driver types and improving their long-term driving behavior – thereby contributing to accident prevention.
This project aims at
– Establishing whether drivers that get feedback regularly drive more safely
– Recognizing and describing different driver types
– Recognizing appropriate feedback types for different driver types
– Evaluating the new feedback system in a real-life context
The project addresses these questions by data analysis, surveys, lab and field experiments .
A new telematics-based personalized feedback system for young drivers that effectively improves their driving behavior.
Digital Stress Interventions
Three out of ten employees in Switzerland experience critical levels of stress at the workplace. Chronic stress can have detrimental effects on our health. The question arises whether new technologies such as artificial intelligence (AI) and virtual reality (VR) can help to detect and manage stress.
To this aim, we are developing a stress detection system based on machine learning and a VR-supported stress management training. The project is conducted in collaboration with international researchers from the fields of psychology, data science and computer science
Stress detection with ML:
Detection of stress levels from data sources available in office environments (i.e., mouse movements,
keystroke dynamics, cardiac activity) with the help of ML.
Stress management with VR:
Heart rate variability biofeedback intervention: Gaining control over stress-related psychological and
physical processes via slow and paced breathing and visualisation of heart activity in VR.
Ethical aspects of digital health:
Investigation of employees’ value-related concerns and wishes for a digital stress management
intervention (dSMI) to shape its development, design and deployment at the workplace.
Stress detection with ML: detection of self-reported stress from mouse movements, keystroke dynamics and cardiac activity collected in a lab experiment (Naegelin et al 2023). Currently, we are validating the results from the lab experiment in a real office environment, developing an ML pipeline capable of detecting self-reported stress from mouse movements, keystroke dynamics and cardiac activity based on real-life office and home office data
Stress management with VR: a single training session resulted in a reduction of perceived stress, an improvement of mood and an increase in feelings of relaxation. Moreover, increases in heart rate variability target parameters were significantly higher in VR than on a classical screen (Weibel et al. 2023). Furthermore, the use of our VR intervention in a 4-week intervention programme lead to improvement of both physiological and psychological measures of stress and stress-related symptoms in compairison to a control group without an intervention (Kerr et al. 2023).
Ethical aspects of digital health: an online study revealed that intention to use a digital stress management intervention at the workplace were moderate to high (Kerr et al. 2023). Employees’ concerns included worries that an intervention would not be effective or even amplify their stress levels. Privacy and accountability concerns were higher if the intervention including artificial intelligence .
Jasmine Kerr, Mara Nägelin, Raphaël Weibel
Holographic Data Science Education
Data science is redefining organizational decision-making in most areas of our lives; yet this transformation is fraught −especially in management − with noteworthy methodological complexity. Traditional data science educational offerings introduce inherent difficulties to students that impede their learning objectives, through
1. the low level of interaction during the learning experience, as well as delayed feedback (this stands in stark contrast to case-based learning),
2. the difficulty in finding an entry point to data science that is properly balanced with their educational background(s), and
3. limited capabilities to reduce complexity through mathematical constructs only.
This project offers a new format for data science education at ETH involving collaborative learning between student bodies with different backgrounds and holographic reality technology. The goal is to improve learning strategies in order to
1. better understand data science methods, including machine learning modeling and decision-making, and
2. analyze & evaluate results with an eye on the trade-off between model performance and interpretability
through a use case-based interactive team experience with 3D immersive technology, i.e. holographic reality.
The project is in collaboration with the Chair of Management Information Systems (MIS) at the ETHZ department of Management, Technology, and Economics (MTEC).
The results show that our collaborative immersive analytics (CIA) system can elicit sustained collaboration among users with different backgrounds. Furthermore, we provide recommendations for the design of CIA systems that enable interdisciplinary teams to jointly solve ML tasks:
Takeaway 1: Pair analytics can be an effective method to elicit collaboration for ML
tasks in co-located, synchronous immersive settings, as it is intuitive and allows users to
share context and visualizations effectively.
Takeaway 2: Clear role division promotes effectiveness in solving ML tasks.
Takeaway 3: Designers of CIA systems should explore ways to encourage critical thinking
in order to avoid overconfidence in ML models.
Takeaway 4: The use of multiple types of interfaces (2D and 3D) does not impede
collaborative efficiency when solving ML tasks.
Takeaway 5: Handovers, although decreasing efficiency, are a driver of collaboration in
co-located, synchronous IA; designs of CIA systems for ML should anticipate their use.
Takeaway 6: Designing CIA systems should consider mechanisms to (re-)establish common
ground, especially for complex ML visualization (e.g., through affordances).
Takeaway 7: Consensus in collaboration and effectiveness in ML modeling within CIA
systems is driven by prior ML knowledge, thus highlighting the need for ML training
among both user pairs.
Takeaway 8: The use of several different interfaces (2D and 3D) does not reduce users’
levels of engagement in solving ML tasks, especially for users who are more knowledgeable
in ML, and may help avoid mental and physical overload.
Ferrario, A., Weibel, R., & Feuerriegel, S. (2020, April). ALEEDSA: Augmented reality for interactive machine learning. In Extended Abstracts of the 2020 CHI Conference on Human Factors in Computing Systems (pp. 1-8).
Benk, M., Weibel, R. P., Feuerriegel, S., & Ferrario, A. (2022). ” Is It My Turn?” Assessing Teamwork and Taskwork in Collaborative Immersive Analytics. Proceedings of the ACM on Human-Computer Interaction, 6(CSCW2), 1-23
Michaela Benk, Raphaël Weibel, Dr. Andrea Ferrario
Trust, Ethics and Algorithms
A wide debate about the ethics of what is commonly denoted as ‘artificial intelligence’ has emerged; it captures the tensions, doubts and fears that are manifested in a greater portion of modern societies when confronted with the conception and use of artificial intelligence-based services and products, in different domains of everyday life.
The plethora of available ethical guidelines from both institutions and companies attemps to be an answer to the urgent questions arising from such interdisciplinary debate. However, the lack of consistency around definitions of artificial intelligence, algorithms and modeling, trust as well as trustworthiness, strongly impacts the quality of scientific inferences and hinders the possibility of consistently and efficiently tackling the aforementioned challenges.
Focus on the fundamentals, i.e.
– trust and trustworthiness: definitions, analysis of accounts and scenarios (human-to-human, e-trust etc.)
– ontology of artificial intelligence and statistical learning algorithms
– the role of transparency and interpretability on artificial intelligence: explanations as communication, the use of counterfactuals in machine learning research
– review of ethical guidelines in relation to artificial intelligence: limitations of top-down approaches (e.g. from a selection fundamental human rights to ethical values and requirements for ethical artificial intelligence), the role of bottom-up contextualized approaches around data-pipelines of data-driven or artificial intelligence-empowered services and products.
The project is run in collaboration with the Institute of Biomedical Ethics and History of Medicine (IBME) at the University of Zurich, the Collegium Helveticum and the Department of Mathematics at the Politecnico di Milano.
Improved understanding of the properties of trust and trustworthiness dynamics as well as explainability and transparency in the presence of selected artificial intelligences in a contextualized setting. Contributions to the debate on the ethics of artificial intelligence by addressing the peculiarities of predictive modeling and algorithms from a philosophical perspective, with the creation of context-driven guidelines.
Relevant publication (selected):
Ferrario, A., Loi, M., & Viganò, E. (2020). In AI we trust incrementally: A multi-layer model of trust to analyze human-artificial intelligence interactions. Philosophy & Technology, 33, 523-539.
Loi, M., Ferrario, A., & Viganò, E. (2021). Transparency as design publicity: explaining and justifying inscrutable algorithms. Ethics and Information Technology, 23(3), 253-263.
Ferrario, A., & Loi, M. (2022). The Robustness of Counterfactual Explanations Over Time. IEEE Access, 10, 82736-82750.
Ferrario, A., & Loi, M. (2022, June). How explainability contributes to trust in AI. In 2022 ACM Conference on Fairness, Accountability, and Transparency (pp. 1457-1466).
Ferrario, A., Gloeckler, S., & Biller-Andorno, N. (2023). Ethics of the algorithmic prediction of goal of care preferences: from theory to practice. Journal of Medical Ethics, 49(3), 165-174.
Dr. Andrea Ferrario
Contextualized Digital Reminiscences
Semantic analysis of multi-scale health data represents a powerful scientific tool to study healthy aging.
The key challenge is to harness the scientific value of multi-scale health data (from both individuals and populations) for a contextualized understanding of health dynamics, including interventions. The latter need to consider the 4 pillars of the healthy aging paradigm, i.e.
– functional abilities
– intrinsic capacities
– decision-making processes of individuals
We focus on reminiscence, i.e. the recollection of significant life experiences, in a healthy aging paradigm by performing data analytics and machine learning (e.g. NLP) on unstructured data to automatically classify reminiscence events and types during everyday life. Automated classification models support the design of digital interventions aimed at the promotion or support of the healthy aging paradigm.
The project is run in collaboration with the Department of Psychology – Gerontopsychology at the University of Zurich and the Collegium Helveticum.
Improved understanding of contextualized reminiscing through data analytics and machine learning and design, deployment as well as testing of digital interventions. Specification of domain specific models of healthy aging, including the clarification of terminology to support interdisciplinary scientific collaboration.
Ferrario, A., Demiray, B., Yordanova, K., Luo, M., & Martin, M. (2020). Social reminiscence in older adults’ everyday conversations: automated detection using natural language processing and machine learning. Journal of medical Internet research, 22(9), e19133.
Stoev, T., Ferrario, A., Demiray, B., Luo, M., Martin, M., & Yordanova, K. (2021). Coping with imbalanced data in the automated detection of reminiscence from everyday life conversations of older adults. IEEE Access, 9, 116540-116551.
Ferrario, A., Luo, M., Polsinelli, A. J., Moseley, S. A., Mehl, M. R., Yordanova, K., … & Demiray, B. (2022). Predicting working memory in healthy older adults using real-life language and social context information: A machine learning approach. JMIR aging, 5(1), e28333.
Dr. Andrea Ferrario
Data Science for Actuaries
The insurance industry is undergoing profound change: data and technology are transforming the insurance business, enabling to management of new sources of risks and the realization of unseen opportunities. New professional players are coming on the market: actuaries now have the opportunity to become active actors of the data and technology transformation or be at risk of slowly being pushed aside by data scientist newcomers.
Augment the actuarial skill set and capabilities through a series of data science tutorials containing both theory on machine learning modelling methodologies and end-to-end insurance use cases. These are supported by scripts in R, Python and H2O.
The project is part of the SAV* Datascience working group initiative.
*Schweizerische Aktuarvereinigung (The Swiss Association of Actuaries)
Publication of case studies to further develop the actuarial skill set by including state-of-the-art machine learning methodologies applied in fully reproducible insurance use cases. Contributions by our Lab include tutorials on Neural Networks (https://dx.doi.org/10.2139/ssrn.3226852), Boosting (https://dx.doi.org/10.2139/ssrn.3402687) and Natural Language Processing (https://dx.doi.org/10.2139/ssrn.3547887).
Dr. Andrea Ferrario
Simulation and prediction of crime is an important pillar of prevention. Due to the increasing digitalization of our daily lives, new data sources are becoming available. The project Home Safety investigates how new data sources and future-oriented methods of analysis can improve the simulation and prediction of crime and, as a result, public safety.
This project aims at providing support in estimating and predicting crime through analysis of the past events and forecasting the potential risk level for the future. The forecasts and simulations are based on historical records. In addition, external data sources are integrated to increase the accuracy of the models.
One part of the research focuses on the prediction of theft and robbery in urban areas. New data such as subway entries and exits, check-ins on social media networks and taxi rides are used to map human activity. Using machine learning, we can predict the areas in a city where crimes can be expected. A new model improves the prediction rate by up to 30% compared to previous models.
Another part of the research simulates the movements of potential criminals in cities. The model uses publicly available data to map the urban structure; it also uses data from a social media network to spatially map human activity in the city. The movements of potential offenders are simulated through agent based modelling. The developed model can in particular serve to test theories as well as a variety of prevention strategies.
Cristina Kadar, Raquel Rosés Brüngger
Small and medium enterprises (SMEs) play an important role in the economy of many countries. This is especially true for Switzerland where SMEs represent more than 99% of all businesses and about 68% of all jobs, acting as the country’s backbone for growth. With the emergence of data mining in the research field of SME revenue and growth, researchers have recently turned their focus to applying data-mining techniques to revenue and growth prediction. However, most of the state-of-the-art prediction models for revenue and growth only include one or few data types such as financial or operational data, which cannot explain the whole and complex context of the modern business environment.
To better understand SMEs’ business performance, we enhance existing business prediction models by leveraging the potential of data mining and web mining for SME growth prediction. Web mining methods are explored with the goal to automatically extract valuable information hidden in the web, whereas data mining methods are applied in order to develop SME growth prediction models with high performance and applicability.
The features derived from the data sources differ significantly in their relevance for the prediction models. The existence of a patent, business sector and age of the entrepreneur were particulary important for the models. The occurrence of SME revenue growth can be predicted using information from the different data sources with approximately 70% of accuracy.
Large amounts of web content are essential for predicting SME growth based on web mining. Only sectors with business areas which are sufficiently covered on the web, such as the hospitality sector, are suitable. Using web mining, the growth of restaurants’ revenues can be predicted with 68% accuracy.
As a secondary result, the project revealed that – on average – SMEs with one or two patents grew 10% faster than comparable SMEs without patents. This growth effect is even stronger at 13.4% for SMEs with inventor teams consisting of both men and women. With respect to the frequency of patent applications the differences between sectors are considerable. Manufacturers of precision devices and electrical elements applied for the most patents.
Dr. Daniel Müller, Dr. Funk Te
The rapid development and penetration of smart and ubiquitous technologies, such as mobile devices and Web 2.0 platforms, have significantly changed the behavior and interactions of consumers. Companies’ offers can be compared more easily. Moreover, consumers want to be able to reach call centers quickly and engage with companies via multiple channels. Yet, according to findings from the World Insurance Report 2013, only 30% of customers report having positive experiences with their insurers, thus clearly indicating that insurance companies should better understand their customers’ needs.
Approach and Results
This project analyzed the interaction between customers and insurers in terms of channels used for interaction in three studies.
The first study analyzed the buying behavior of insurance customers in a multichannel distribution system. The analysis compared customers who bought their insurance policy directly online on the website (Group 1) with customers who first obtained information on the price and product on the website, but then bought their insurance policy through the agency (so-called RoPo customers, Group 2), and customers that were solely in contact with the agency during the purchasing process (Group 3). The study suggests that customers in Group 1 and 2 are younger, more likely to live in urban environments and more likely to be new clients. Moreover, customers of Group 2 had the highest average insurance premium. The average premium volume of these customers was over 5% higher than for those in Group 1 for two out of three insurance products examined.
The second study focused in particular on the buying behavior of RoPo customers (research online, purchase offline). The aim was to understand how much time these customers need to make the decision to purchase. It was shown that RoPo customers took an average of 19 working days from researching the product online to buying it at the agency. This timespan varies depending on the type of insurance product. It is lowest for travel insurance with 15 working days, while customers need an average of 20 working days to decide on household/personal liability and motor insurance policies.
The third study aimed to predict cancellations and renewals of existing policies as well as purchases of additional products (cross-buying). It was shown that it is possible to make an accurate prediction of the future buying behavior of RoPo customers based on their linked internet data. The error rate of the models to predict cancellations and purchases of additional products was as low as 20% to 25%.
Dr. Stefan Mau