Exact forecasts in the company are the basis for a useful, proactive acting and planning. Interestingly, the weighted average of not fully correlated single predictions or expectations usually provides better results than the best individual forecast or expectation. Even though it is common knowledge that combinations are usually advantageous (and under certain conditions it should be proved to be better), no analytical decision models exist comprising which individual forecasts should be combined and how they should be weighted relatively to each other. In our research, we develop asymptotically optimal weighting schemes of forecasts by means of new models of the statistical learning theory in order to determine and minimize the error components of combination methods.

Recommender systems have the task to predict the interest or the benefit for a user regarding an object in order to recommend the most suitable and interesting objects to this specific user. The typical area of application comprises the products of a web shop, music tracks, movies, web sites shown by search engines and personalized advertising banners in the internet, especially in affiliate networks. In order to detect the suitable recommendations, methods of machine learning and information retrieval are used. In our research we develop and test new statistical-mathematical methodologies for the hybridization of recommender systems, in other words for the profitable integration of different recommender systems (different algorithms that are based on different data) in order to improve the predictive power.

Complex operational planning and optimization problems often have a large number of constraints, so-called dimensions, where the computation and storing complexity usually increases super linearly with the corresponding dimensionality. Thereby many practical problems cannot be solved optimally anymore and solution heuristics of operations research are used. In our research we therefore take new geometric approaches, where the matrix of the constraints can be described by a low dimensional matrix by means of empirical-orthogonal decomposition in such a way that new optimization problems can be formulated, which are also able to solve the original problem, however having a much lower processing complexity.

Nowadays decisions are increasingly made in an automated or data-supported fashion by means of decision support systems (DDS). Although DDS aim to support as rational as possible and undistorted decisions, empiricism shows that even if decisions are made with modern DSS, they are still affected by the so-called cognitive biases of the user of the DSS and consequently are suboptimal. In our research we work on new forms of DSS that extract typical distortion patterns of the decisions made in the past in a data analytical way and make them available to the decision maker together with automatically generated explanations (explanation engineering) as information objects for the purpose of self-reflection. The project idea is currently being piloted by a big German company.

Churn prevention campaigns aim to identify customers early that are at risk of churning by means of data analytical methodology and to prevent churn with appropriate actions before termination. Therefore, not only churn probabilities but also the reduction potential per customer have to be estimated as accurately as possible, which in practice is not achieved sufficiently for various reasons. Such campaigns often even increase the termination rate and are used only rarely. In our research, we therefore develop more robust methods of probability difference estimation, among other things based on distance-searching pruning of differential classification trees and hierarchical bootstrapping approaches.

Nowadays, industrial plants and vehicles feature a large number of sensors that record the behaviour as well as the state of technical components. Predictive maintenance aims to anticipate critical states as early as possible and act appropriately (proactive exchange of components, adjustment of operating modes, finding a workshop etc.) by means of utilizing data that is made available in data warehouses. In particular, the sequential or chronological progress of the development of different sensor values are of special interest in order to derive trends, correlations and causal-diagnostic statements. In our research we therefore develop methods of the "frequent sequential pattern mining" in multivariate field data. Based on that, the methodology of machine learning and network modeling is used in order to determine the probability distribution of critical events over time and consequently being able to act rationally and economic pro-actively.