Metalearning

One of the challenges faced by data mining practitioners is the selection of the most accurate algorithm for their classification tasks. Given that each algorithm performs well only on a subset of classification tasks --a direct consequence of the No Free Lunch theorems (Schaffer, 1994; Wolpert, 2001), and that there is a growing number of available algorithms, finding the best algorithm for a particular classification task is indeed becoming increasingly difficult.

For the practitioner, what is needed is an automatic system capable of returning the most suitable algorithm for his/her task. Meta-learning, or the use of data collected from the application of data mining to build metamodels that map classification tasks to algorithms, has proven a viable solution for the design and implementation of such systems.

Our research in meta-learning focuses on gaining insights on the mechanisms of learning, including dataset characterization and algorithm behavior, and using these in the design of decision-support systems for practitioners.


Click here for some of our publications, and here for our most recent Springer book on metalearning. Interested lecturers are encouraged to use the free "Online Examination Copy" option on the bottom-right of the page.

We also have a Google Group on Meta-learning. Feel free to visit and contribute.