Partenaires et organisations internationales
(Anglais)
|
University of Bristol (UK), DaimlerChrysler (D), ISL-SPSS (UK), Austrian Institute for Artificial Intelligence (A), University of Porto (P)
|
Résumé des résultats (Abstract)
(Anglais)
|
The METAL project aims at the development of methods and tools for providing support to users of machine learning (ML) and data mining (DM) technology. While the interest in such technology, particularly in the area of classification and prediction, is growing rapidly in industry and commerce, and a number of data mining tools are already available, such tools are still of limited use to end-users who are not experts in machine learning. In reaction to this situation, the central goal of METAL is to develop a prototype assistant system that supports users with model selection and method combination, and guides them through the space of experiments.
Achievements so far :
· Collection and characterization of three classes of meta-objects manipulated in METAL: (1) applications and datasets, (2) learning/data mining algorithms, and (3) data pre-processing tasks and methods.
· Extensive experimentation in view of meta-data collection for meta-learning. Ten symbolic and numerical data mining algorithms were selected and applied to more than 100 datasets for classification tasks, and a dozen other algorithms were tested on around 50 datasets for regression tasks. In all, several dozen thousands of experiments were conducted, and the results are stored in a repository of empirical meta-data for meta-learning.
· Design, implementation, and comparative evaluation of a number of novel meta-learning strategies in view of model selection (e.g., data characterization tool based meta-learning, landmarking, histograms, ranking, zooming), all of which have been presented in international conferences and journals.
· Construction of a large, online, dynamic, annotated bibliography of theoretical and experimental work on meta-learning.
· Specification, design, and implementation of several ranking strategies (e.g., based on data envelopment analysis or on adjusted ratio of ratios) that can combine multiple evaluation criteria such as predictive accuracy and computational efficiency.
All the above achievements have been integrated into a prototype Data Mining Advisor which is available for online consultation via the Web.
|
