A Case-Based Meta-Learning and Reasoning Framework for Classifiers Selection

In machine learning area, a large number of classification algorithms are available that can be used for solving the problems of prediction and classification in different domains. These classifiers perform differently on different learning problems. For example, if one algorithm perform better on one dataset, the same algorithm may perform badly on another dataset. The reason is that each dataset has its own nature in terms of its local and global characteristics. Similarly, the number of candidate algorithms are also large in number and is therefore very hard for a machine learning practitioner to know the intrinsic behaviors of the algorithms on different kinds of datasets and are therefore unable to select a right algorithm for his problem in-hand. To overcome the issue, this study proposes an automatic classifier selection methodology. A case-based meta-learning and reasoning (CB-MLR) framework is designed and implemented to recommend appropriate classifier for mining the new dataset. The framework exploits inherit characteristics of the datasets mapped against the algorithms performance. The key contributions of CB-MLR include: (a) design of a flexible and incremental meta-learning and reasoning framework using multi-view learning, and (b) implementation of the CBR methodology to accurately recommend most relevant top-3 classifiers as the suggested algorithms for the new data mining problem. The proposed framework is tested for 9 decision tree classifiers, from Weka environment, and 52 datasets from UCI repository over a case-base of 100 resolved cases. The accuracy obtained is 94% within the scope of top-3 most relevant classifiers.