Our General Research Objective - The research at DSAR focuses on the development of robust and fair machine learning methods with high predictive quality and interpretability. Currently, this involves the following considerations:

• The generation/selection of robust feature sets for complex (stream-) learning procedures
• Mechanisms for robust and fair predictions
• Techniques for the quantification and improvement of data quality

In general, we aim to provide the groundwork for feasible machine learning solutions in various practical applications.

Bayesian unsupervised ensemble methods in Distributed learning: In this work, we develop Bayesian aggregations on portioned data. Local estimators (or experts) are trained in different subsets of data and due to the divide-and-conquer approach,  they should be aggregated. Different scenarios are available, e.g. Product of Experts, Bayesian Committee Machine, and so on. Since they assume independence between local experts,  in practice they lead to a poor prediction. We want to develop an aggregation using correlated experts and Bayesian unsupervised ensemble technique in Random /disjoint partitioning by a focus on Gaussian process.

Graph-based aggregation in Local approximation: When we train a set of experts on the local subsets of training data, the interaction between experts can be used to develop a fast, accurate, and consistent predictive distribution. Since the state-of-the-art method which uses the correlation between the experts is not efficient for the large and high-dimensional data sets, we use graphical model and conditional independence to choose the most important experts and connections for estimation.

Undirected graphical models with latent variable in distributed learning:

In this line of work, we consider a graphical model with its observed random variables and a latent variable. The observed variables show the local experts while latent variable is the desired estimator. This work tries to find appropriate joint distribution between the nodes while the aggregation is based on an energy function.

Actionable Recourse Through Data Supported Counterfactuals - Counterfactual explanations can be obtained by identifying small changes made to an input vector to influence predictions; for example, from ’loan rejected’ to ’awarded’ or from ’high risk of cardiovascular disease’ to ’low risk’. Previous approaches often emphasized that counterfactuals should be easily interpretable to humans, motivating solutions with few changes to the input vectors. However, some approaches would not guarantee that the produced counterfactuals be (1) proximate (i.e., close to the current state), (2) connected to regions with substantial data density (i.e., supported by data) and (3) attainable without excessive efforts, three requirements that are fundamental when making suggestions to individuals that are indeed attainable. In our WWW 2020, we use data density approximators (for example VAEs) to find close and attainable counterfactuals.

Relating Data Supported & Greedy Counterfactuals under Predictive Muliplicity - Recent work has revitalized an old insight from the late Leo Breiman: there often does not exist one superior solution to a prediction problem with respect to commonly used measures of interest (e.g. error rate). In fact, often multiple different classifiers give almost equal solutions, which is called predictive multiplicity. In this work, we derive upper bounds for the costs of counterfactual explanations under predictive multiplicity.

Auditing Black Box Classifiers for Proxy Influence -  In this line of work, we develop a method to generate transparent predictions. Those differ from explainable predictions. Our goal is to answer which inputs are being used locally & globally, and to what extent even if they are not used to train the model?

Explainable Deep Learning - We proposed a new layer - CancelOut that can help identify a subset of relevant input features for streaming or static data; the code is available online.

Multi-Output Regression - We open-sourced a python framework dedicated to multi-output regression problems - amorf. 

Deep Learning for Tabular data - Nowadays, artificial neural networks have become the main tool for machine learning tasks within a wide range of domains, including vision, NLP, and speech. However, deep neural networks show moderate performance on heterogeneous data. We would like to address this issue. 

Seite im Aufbau...