|
NeuroCOLT
Technical Report NC-TR-97-025
On
Learning from Multi-Instance Examples:
Empirical Evaluation of a Theoretical Approach
Peter
Auer
Technische Universitaet Graz, Austria
Abstract
We describe a practical algorithm for learning axis-parallel
high-dimensional boxes from multi-instance examples. The first solution
to this practical learning problem arising in drug design was given
by Dietterich, Lathrop, and Lozano-Perez. A theoretical analysis was
performed by Auer, Long, Srinivasan, and Tan. In this work we derive
a competitive algorithm from theoretical considerations which is completely
different from the approach taken by Dietterich et. al. Our algorithm
uses for learning only simple statistics of the training data and
avoids potentially hard computational problems which were solved by
heuristics by Dietterich et. al. In empirical experiments our algorithm
performs quite well although it does not reach the performance of
the fine-tuned algorithm of Dietterich et. al. We conjecture that
our approach can be fruitfully applied also to other learning problems
where certain statistical assumptions are satisfied.
Download Compressed Postscript
|