The Machine Learning Group at the Leuphana University of Lüneburg has been established in 2015.
The Web has become an everyday resource and simple queries, such as the search for an address of a restaurant, are efficiently processed by look-up tables and search engines. By contrast, searching the Web becomes tedious when the information being sought is distributed across distinct and heterogeneous sources; the user herself needs to find and aggregate trusted sites that contain relevant information.
However, content that serves one user does not necessarily satisfy the information need of another as users exhibit different opinions about trusted sites, types of media, and finally also relevance. In this project, we study the personalized acquisition of content from heterogeneous sources. We aim at aggregating relevant content for a given query from news articles, blogs, tweets, discussion forums, etc. The goal is not only to assemble the big picture but also to incorporate opinions and sentiments in the summary.
In cooperation with the German Institute for Educational Research (DIPF) we work on machine learning for educational research. Educational research deals with all facets of education and its assessment. Perhaps the most prominent assessment is the international PISA study launched by the OECD in 1997. The triennial study evaluates education systems in about 70 countries in the subjects reading, mathematics, and science. We focus on computer-based assessments and study adaptive testing to personalize tests in realtime (tailored testing).
We phrase tailored testing as a learning problem and study novel machine learning methods to personalize the selection of test items to advance the state-of-the-art in adaptive testing. Furthermore, we are interested in general data mining and analysis problems in educational research to better understand the underlying processes. Moreover, educational research bears many open problems and interesting research questions from a machine learning point-of-view. Besides computer-based assessment, we are interested in general data mining and data analysis problems for educational research such as human planning strategies.
In cooperation with Zalando SE we study large-scale recommender systems and personalization techniques. We investigate novel approaches to on- and off-site recommendations and ways to personalize web layouts to support user navigation and to improve user engagement.
Recommender systems help users to navigate through web sites and to highlight content that may not have been found otherwise. In commercial scenarios, their performance is directly connected to business strategies and to the overall revenue of the company. Amazon and Netflix are prominent examples for a large-scale deployment of recommender systems. Data is being collected in form of user feedback, which is often given in form of ratings from 1 (dislike) to 5 (like), purchases, and page views or clicks. Collaborative recommender systems mine the user feedback to identify similar users; recommendations are then computed from this neighborhood. Besides relevance, frequently desired properties of the recommended items are for instance freshness and diversity.
The formal problem setting of recommender systems matches that of reinforcement learning. Exploration and exploitation are thus common principles to select recommended items from the pool that usually follows a power law. Recommending items from the long tail however remains challenging. Personalization approaches could remedy the uncertainty by learning a model for every user. However, infrequent users hardly leave enough feedback to be captured well and alternatives such as hybrid models need to be deployed.