When: Tuesday, June 3
Time: 1 pm  (please note unusual time and day)
Where: CS 2310 
Speaker: Filip Zelezny, Post-Doc in the Dept. of Biostatistics & Medical Informatics

Title: Propositionalization via First-Order Logic Feature Construction

One way to solve machine learning problems on multi-relational/structured
data is to transform the data into a flattened, single-relational
representation and then take advantage of a broad variety of learning
algorithms available for this simpler ("propositionalized") form of data.
The RSD system solves the transformation task by constructing first-order
logic features. It incorporates several heuristically inspired pruning
rules as well as ways of naturally bounding the feature-language, to
prevent the threat of combinatorial explosion intrinsic to the
transformation task. The system can export the flattened data into
formats directly acceptable by the systems Weka, CN2 and others.  In the
seminar, I will review the employed principles and related work briefly,
and then switch to a demonstration of RSD applied to some classical
challenges of multi-relational learning. RSD is publicly available from
http://labe.felk.cvut.cz/~zelezny/rsd and it comes with a comprehensive
user's manual.  (The website also mentions some data-mining capabilities
of the system, which I will not cover in the talk, unless prompted to).