Trustworthy AI for Seamless Problem Solving: Next Generation Intelligence Joins Robust Data Analysis

RTG DataNinja

A small letter n with a stylised black ninja mask

X-FI - Explainable Feature Importance: Interpretable Machine Learning through Game-Theoretic Analysis of Influencing Factors and Interaction Effects

This project aims at transparent modeling of dependencies and interactions between variables as well as theoretically justified reductions of nonlinear dependencies to understandable and interpretable measures for the influence of individual variables, both for supervised and unsupervised learning. A theoretical framework of that kind shall be complemented by suitable visualization techniques to further facilitate interpretation, the interaction with users and domain experts, and the exploration of dependencies in high-dimensional data spaces. All methods shall be evaluated in various case studies in the medical domain.

The main innovation of the project is to establish a connection between feature analysis and (cooperative) game theory. To this end, a feature subset is considered as a coalition in game theory, and the performance of a model trained with these features as the value of the coalition. The importance of individual variables can then be quantified in terms of well-established game-theoretical measures, such as the Shapley value, just like the interaction between pairs or groups of variables. Due to the exponential size of the lattice of feature subsets, the exact computation of such quantities is in general infeasible. Therefore, a key challenge of the project is the development of efficient approximate algorithms.

Joint project with, Prof. E. Müller (TU Dortmund)

EKAmBa: Realtime Configuration of Algorithms with Multi-armed Bandits

The parameters of an algorithm have an influence on its problem-solving behavior and hence its performance on specific problem instances. For example, depending on how skillful its parameters are set, an algorithm for solving traveling salesperson (TSP) problems may run faster or slower on a concrete TSP instance or may produce better or worse solutions (an approximation to the optimum). Algorithm configuration (AC) refers to the task of automatically adjusting the parameters of an algorithm on a per-instance basis, i.e., tailoring the parameters to a specific problem instance​. We are especially interested in a practically motivated pool-based scenario, where problems of the same type are solved repeatedly, and several algorithms are run on each problem instance in parallel. In this context, parameters ought to be adjusted in real-time.

To tackle the AC problem, we make use of multi-armed bandit algorithms, a class of machine learning methods for sequential decision making under uncertainty​. More concretely, we will develop new types of MAB algorithms specifically tailored to our setting, which are able to choose high-quality parameters even as the underlying distribution of problems drifts over time. To this end, we will mainly build on so-called preselection bandits​​, a recent class of bandit algorithms that is well suitable to formalize the pool-based variant of real-time AC​.

Joint project with Prof. K. Tierney (Bielefeld University)

Funding
RTG Dataninja (Trustworthy AI for Seamless Problem Solving: Next Generation Intelligence Joins Robust Data Analysis), State North Rhine-Westphalia
Duration
11/2020 - 11/2024
Role
Joint project with, Prof. E. Müller (TU Dortmund)
Project Website
X-FI
EKAmBa