Learning With Monotonicity Constraints
Stc
Date: 2011-05-13
Time: 11:00 - 12:00
Room:
BBL 065
Speaker: Ad Feelders
Title: Learning with Monotonicity Constraints
Abstract
In many applications of data mining one knows beforehand that the target should be increasing (or decreasing) in one or more input attributes. Consider for example the classification of documents with respect to their relevance to a particular query. We could classify documents as {not relevant,somewhat relevant, relevant} based on attributes such as the number of query terms that occur in the abstract or title. Common sense dictates that relevance should be increasing in these attributes.
We start with a general introduction to data mining with such so-called monotonicity constraints. Then we discuss in more detail an algorithm to make data sets monotone while minimizing a given convex loss function such as squared error.
