University of Helsinki Department of Computer Science

Department of Computer Science

Department information


58308109 Seminar: Predicting Structured Data (Spring 2008)

Time: periods III-IV, Thursdays 16-18, room C220
Organizers: PhD Huizhen Janey Yu , prof. Juho Rousu


Complex learning targets such as sequences, taxonomies and graphs are frequent in real-world applications, for example, hand-writing recognition (target is a sequence of letters), hierarchical classification (tree), and gene function prediction (poset), to only name a few.

The mainstream of machine learning research, in particular that of kernels methods, has been succesful in developing flexible and powerful methods for treating complex inputs. The complementary methods for complex outputs have so far received significantly less attention. The chief approachs towards complex targets has been to decompose the target (e.g. a hand-written word) prior to learning and learning each component (e.g. a character) indepedently. With this approach, dependencies between the components are not utilized.

During last five years, research in complex and structured output learning has emerged as one of the mega-trends in machine learning. In particular, methods marrying kernel methods and graphical models have received significant attention.


The purpose of the seminar is to explore the recent progress in machine learning for complex and structured outputs


The seminar is an elective advanced level seminar. It is also well-suited for post-graduate studies.

Prequisite knowledge for the semimar is basic knowledge about probabilistic modelling and machine learning. Familiarity with kernel methods and graphical models will be helpful.



Kernels for Structured Data

String kernels

Rational kernels

Tree kernels

Graph kernels

Structured Prediction Models

Structured Prediction Applications

Sequence annotation

Hierarchical multilabel classification

Supervised network inference & completion

Optimization algorithms for structured prediction

Generalization error analysis for structured output