wir bieten...
Dekobild im Seitenkopf ISMLL
The Information Systems and Machine Learning Lab (ISMLL) is devoted to research in machine learning and data mining. Our main expertise is the development of machine learning models and respective learning and inference algorithms for supervised problems with complex data, i.e., data that cannot be easily described as a set of instances with a fixed set of attributes. Instead of developing isolated application-specific solutions, we aim at mapping application problems to abstract problem classes, formally described by some common characteristics, and then generically tackle these problem classes. Especially, we work in the following areas:
  1. Factorization methods address problems with categorical variables with many levels (as e.g., in recommender systems) by associating latent features with every level. Interactions between these variables then can be captured in models by interactions between the latent features. Two-way interactions lead to matrix factorization models, higher interactions to tensor factorization models.
  2. Relational classification aims to improve classifiers by relational covariate information, i.e., further relations besides the target relation. For example, if publications should be classified by their topic, the author, venue and citation relations will contain crucial information.
  3. Time series classification is the task to assign time-variant data to predefined classes, e.g., to recognize if sensor data such as EEG indicates an healthy or an ill condition of a patient. Non-aligned time series can be utilized by different means such as kernel methods or motif extraction.

Furthermore, we work on some related problems and complementary aspects such as record linkage, frequent pattern mining, active learning, and ranking.

In the scope of research projects, industrial cooperations and lab experiments, our methods support various applications:

  1. Recommender systems are dynamic adaptive systems for personalization that help users to chose between alternatives and to find items they are interested in by learning their preferences from collective past user behavior. We applied such systems
    • in online shops in e-commerce (LEFOS)
    • for recommending tags in web information systems (ECML/PKDD 2009 challenge)
    • in IPTV and web-based Multimedia applications (MyMedia), as well as
    • in museums (REMIX)
  2. Object recognition in high-volume spatio-temporal sensor data engineering applications, e.g. for vibration analysis and mode line detection (X-Media),
  3. Knowledge extraction from text and other un- or semistructured sources into formal representations such as ontologies in the context of the semantic web, e.g. for
    • extraction of relation instances from wikipedia and
    • event detection in tagged photo collections.
  4. Adaptive tutoring systems in e-learning track the progress of learners while solving exercises and aim to chose further exercises to balance success and challenge for the learners.