1. Institut für Softwarewissenschaft - Universität WienP.Brezany 1 Meta-Learning in Distributed Datamining Systems Peter Brezany Institut für Softwarewissenschaft Universität Wien Tel. : 01/4277 38825 E-mail : brezany@par.univie.ac.at Sprechstunde: Dienstag 11.30-12.30

2. Institut für Softwarewissenschaft - Universität WienP.Brezany 2 Introduction Meta-learning (learning from learned knowledge) – a technique dealing with the problem of computing a „global“ classifier from large and inherently distributed databases. A number of independent classifiers – „base classifiers“ -are computed in parallel. The base classifiers are then collected and combined to a „meta-classifier“ by another learning process. The predictive accuracy of base classifiers is improved. Assuming that a system consists of several databases interconnected through an intranet or internet, the goal is to provide the means for each data site to utilize its own local data and, at the same time, benefit from the data that is available at other data sites without transfering or directly accessing that data. The above concept can be materialized by learning agents that can execute at remote data sites and generate classifier agents that can subsequently be transfered among the sites.

3. Institut für Softwarewissenschaft - Universität WienP.Brezany 3 Meta-Learning Scenario Learning Algorithm Training Data Training Data Learning Algorithm Classifier Validation Data Predictions Meta-Level Training Data Meta-Learning Algorithm Final Classifier 1 1 112 2 2 2 3 3 3 4 4

4. Institut für Softwarewissenschaft - Universität WienP.Brezany 4 Meta-Learning Scenario (2) 1.The classifiers (base classifiers) are trained from the initial (base-level) training sets. 2.Predictions are generatedby the learned classifiers on a separate validation set. 3.A meta-level training set is composed from the validation set and the predictions generated by the classifiers on the validation set. 4.The final classifier (meta-classifier) is trained from the meta-level training set.

5. Institut für Softwarewissenschaft - Universität WienP.Brezany 5 Strategies for Combining Multiple Predictions from Base Classifiers 1.Voting - Each classifier gets 1 vote; the majority wins. 2.Arbitration – A prediction of an „objective“ judge (a classifier) is selected if the participating classifiers cannot reach a consensus decision. 3.Combining – the use of knowledge about how classifiers behave with respect to each other. 1.A combiner is a program generated by a learning algorithm that is trained on the predictions produced by a set of base classifiers on raw data (hierarchical structure is possible).

6. Institut für Softwarewissenschaft - Universität WienP.Brezany 6 Example: A Combiner with 2 Classifiers Classifier 1 Classifier 2 Combiner Instance Prediction 1 Prediction 2

7. Institut für Softwarewissenschaft - Universität WienP.Brezany 7 Notation x – an instance (sample) whose classification we seek C 1 (x), C 2 (x),..., C k (x) – predicted classifications of x from the k base classifiers, C i, i = 1, 2,..., k class(x) – correct classification of x attrvec(x) – attribute vector of x E – validation set of examples; x  E T – set of „meta-level training examples“

8. Institut für Softwarewissenschaft - Universität WienP.Brezany 8 Combiner Strategy Combiner = meta-classifier (generated by meta-learner) Composition rule – determines the content of training examples for the meta-learner; it varies in different schemes. 2 schemes for the composition rule, according to the strategy used for computation of T. –Class-combiner. The meta-level training instances consist of the correct classification and the predictions; i.e., T = {class(x), C 1 (x), C 2 (x),..., C k (x) | x  E} –Class-attribute-combiner. T = {class(x), C 1 (x), C 2 (x),..., C k (x), attrvec(x) | x  E} These composition rules are also used in a similar manner during classification after a combiner has been computed. Given an instance whose classification is sought, we first compute the classifications predicted by each of the base classifiers. The composition rule is then applied to generate a single meta-level test instance, which is then classified by the combiner to produce the final predicted class of the original test datum.

9. Institut für Softwarewissenschaft - Universität WienP.Brezany 9 Combiner Strategy (2) Validation Set

10. Institut für Softwarewissenschaft - Universität WienP.Brezany 10 A Real Medical Application