Our research is inspired by the challenges of changing environments. We develop generic methods for model adaption and for change monitoring, and dedicated methods in medical mining and in business mining.
In the thematic area Medical Mining we develop mining methods for the analysis of cohort data from epidemiological and clinical studies. Together with the University Medicine Greifswald we work on learning classification models, on the identification of subpopulations with increased disease prevalence, on the characterization of such subpopulations, and on reducing the high-dimensional feature space in a semi- supervised way. Together with the Diabetology of the OVGU University Medicine, we work on analysing stance patterns for patients with diabetic foot syndrome. Together with the University Hospital Regensburg and the University of Ulm we study the evolution of patients suffering from tinnitus. More on our cooperations and initiatives on Medical Mining can be found here.
In the thematic Area Business Mining we develop stream mining methods for polarity classification and topic monitoring on opinionated document streams and for dynamic recommendation engines. Opinion stream mining is the subject of the project . Learning on a stream of ratings and adaptation of the recommender core is investigated under Dynamic Recommenders.
In the project OSCAR we build stream mining methods to capture the evolution of a stream of opinionated documents. Building upon the results of the earlier project IMPRINT, we strive to develop supervised and semi-supervised (passive and active) learning algorithms that can exploit inputs from the human expert, as well as historical data on the words and entities observed in the stream. More on OSCAR is here.
Model learning and adaption is crucially influenced by the availability of information: the labels needed for supervised learning may be scarce and expensive to produce or may become available with a temporal delay. Under Drift Mining, we work on methods for drift modeling and change monitoring, subject to limitations of information availability; more information here (German only). Under Probabilistic Active Learning we work on learning methods that actively ask for labels - in interaction with the human expert; more information on PAL.
In Recommender Systems we deal with big and sparse data. For building predictive models on this kind of data we use matrix factorization methods together with stream mining techniques. Our focus lies on adaptive methods that are able to take evolution of users' preferences and changes of the environment into account.
Subject of the earlier project IMPRINT was the study of multi-relational objects, which were fed by streams of observations. An example of such objects are patients, for whom medical assessments are recorded in regular or irregular time intervals; one of our application areas in IMPRINT was medical mining. Another example of such objects are products, for which opinions are continuously uploaded; in IMPRINT, we considered opinion stream mining as a further application area. IMPRINT encompassed methods for multi-relational stream classification, unsupervised and semi-supervised stream clustering, methods modeling the evolution of individuals and methods capturing the evolution of an underlying concept (e.g. disease progression). More information on IMPRINT is here.
A detailed list of current and completed projects can be found on the reseach portal of Saxony-Anhalt, click here (in German).