[Solomonov Seminar] 167. Solomonov seminar

[Marko Grobelnik]

Vabim vas na 167. Solomonov seminar, ki bo v torek 24. januarja 2006
ob 13:00 uri v Veliki predavalnici IJS.

Tokrat bomo gostili Luisa Torga iz Univerze v Portu na Portugalskem,
ki bo predaval o motodah za napovedovanje redkih dogodkov.

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Luis Torgo, University of Porto

     Non-Uniform Cost Surfaces for Predicting Rare Extreme Values

In this presentation we will address the problem of rare extreme values 
prediction. Modeling extreme data is very important in several application 
domains, like for instance finance, meteorology, ecology, etc.. Our target 
applications have as main objective to be able to anticipate extreme values 
of a continuous variable. The main distinguishing feature of these 
applications resides on the fact that these values are rare. Any prediction 
model is obtained by some sort of search process guided by a pre-specified 
evaluation criterion. In this work we argue against the use of standard 
criteria for evaluating regression models in the context of our target 
applications.  We describe a new predictive performance metric for this class 
of problems that our experiments show to perform better in distinguishing 
models that are more accurate at rare extreme values. This new evaluation 
metric can be used as the basis for developing better models in terms of rare 
extreme values prediction. We also describe REC surfaces and REC maps that 
are tools that can be used to visualize/compare the performance of regression 
models in applications with non-uniform error costs.  These tools can be seen 
as generalizations of REC curves (Bi & Bennet, ICML03) that introduce a 
further degree of detail by plotting the cumulative distribution function of 
the errors across the distribution of the target variable, i.e. the joint 
cumulative distribution function of the errors and the target variable. This 
provides a more detailed analysis of the performance of models when compared 
to REC curves. This extra detail is particularly relevant in applications 
with non-uniform error costs, where it is important to study the performance 
of models for specific ranges of the target variable.