Know Center Research

Fair AI

What dangers of discrimination in terms of skin color, gender or popularity lie in the algorithms of an AI? What do we ourselves have to plan for in time, what must the AI itself be able to do, and how can we support people in making decisions in a considered and critical way on the basis of AI insights?

Fair AI

Research Focus

The mission of the Fair AI Research Area is to research and develop fair, bias-free algorithms and evaluation methods that minimize the risks of discrimination and promote trust in AI. In close collaboration with the other Research Areas, we are working on the profound understanding of causal relationships underlying wrong decisions. Our goal is to develop fair algorithms and evaluation methods that are essential building blocks of trustworthy AI. We want to support users in their self-determined, critical, and informed decision-making and interaction with AI-based systems (e.g., recommender systems).
Biases in Data and Systems

Cognitive and social biases in human perception and behavior are captured, reflected, and potentially amplified by AI-based systems. Common examples include the promotion of popularity biases and polarization in online information, or the adoption of gender biases in search engines. Our research aims to understand such cognitive and social biases in user interactions and other forms of AI training data.

Fairness in Algorithmic Decision Support

To create the highest possible transparency in our partners’ systems, we explore the reasons for social and cognitive biases in algorithms and develop specific modeling techniques for individual (groups of) users. This will contribute to an ecosystem in which a high level of transparency and fairness is achieved through personalized algorithms.

Privacy-aware Recommendations

Recommender systems typically require a lot of personal and private data to calculate personalized recommendations. We develop privacy concepts for recommender systems and give users control over how much data they share. With highly innovative “obfuscation techniques” evaluated with our partners in sensitive environments like healthcare, we balance data accuracy and privacy.

Potential

Fair AI builds trust.

A perfect AI application, tailored exactly to your business, is good. But it is even better if the AI is also fair. Because then all your customers (or even the users of public services) will trust your company in the long term. And if data is the most valuable resource a company can have today, then trust is the most important resource that can be given to you from the outside. It's better not to put it at risk.
Recommender systems

Personalized recommendations and decision support that can self-optimize based on user feedback

Bias Detection

Are there certain customers/user groups that are treated unfairly?

Bias Mitigation

How can the effects of existing biases be minimized? How can we improve the quality of algorithms?

Certification

Do your datasets and algorithms comply with the new EU AI law? How can they be certified?

Fairness Check

Which fairness criteria should be applied? Is there a trade-off between different definitions of fairness? What about the balance of fairness between providers and consumers? The Fairness Check identifies the status of fairness of an existing application and enables us to find ways to more fairness and trust quickly and efficiently.

User understanding

Cluster user groups based on their interests and understand the connection between user:groups. Applied market research at the push of a button, in real time and perfect for supporting strategic or short-term decisions in product development, sales, or marketing.

Churn Prevention

Identify those users who may leave the system due to their behavior. On this basis, you can develop appropriate individual retention offers or gather valuable feedback. Don’t let customers leave without noticing!

Identify niches

Often, user groups exist that prefer content that does not correspond to the majority. Always identify such niche groups, which are also very likely to have problems finding content that suits them. In this way, you can learn to further diversify your offers or communicate separate offers for these customer groups. The importance of customers who prefer niche offers should not be underestimated.

Provider Fairness

Who brings data into the system? Are all providers treated fairly? Does all content have a chance of being discovered? Increasingly, sharing economy platforms or cluster platforms are where companies and individuals independently provide information, exchange services and products, or transact purchases and sales. Make sure that such a platform treats all providers fairly and thus give the platform the chance to scale!

Consumer Fairness

Who are your customers/users? Do they get good results from AI?

Fair Matchmaking

How can we realize matchmaking between different offers and demands in a system in a fair and transparent way? What are the criteria according to which matchmaking should work? The goal is to achieve the best result in the end among a multitude of possibilities.

Data Quality Assesment

What is the quality of the training data in the system?

Long-Term Simulations

Analyze the impact of privacy-preserving technologies on the precision of decision support and recommender systems, and evaluate privacy-preserving data.

Anonymous Users

Session-based recommendations for anonymous users who are not logged in, using deep learning methods. Many users would rather not log in, especially for not so large portals, or are only quickly seeking information or products. From an economic perspective, however, these users should also be able to find the same information or store for the same products as registered users in most cases.

Protect privacy

The use of private data is often not permitted. We have to make sure that AI systems cannot draw any conclusions about users based on the data collected during the session.

User Control

Allow users to decide from the outset how much and what data they want to share with the system by designing the application. But also offer the option to change this again. Transparently, without any fine print.

Trade-off between precision & privacy

Analysis of the impact of privacy-preserving technologies on the precision of decision support and recommender systems and the evaluation of privacy-preserving data.

Consulting

Advice on recommendation algorithms and data sources, and evaluation metrics.

Dynamic Pricing

Dynamic adjustment of the price based on various criteria, e.g., fairness criteria or for optimizing turnover

Personalized Search

Combination of search technology with recommender systems. Customized rankings for different user groups using, e.g., Learning2Rank.