Algorithm Selection based on Landmarking Meta-feature

Abstract

Knowledge discovery is the data mining task. Number of classification algorithms is present for knowledge discovery task in data mining. Each algorithm is differentiating with another based on their performance. No free lunch theorem [1] states that there no single prediction of algorithm is not possible for all kind of datasets. This implies that performance value of algorithm changes according to dataset characteristics. Non-expert can't understand which will be best classifier for his/her dataset. Meta-learning is one machine learning technique which supports non-expert users for selecting classifier. In meta learning dataset characteristics well know as meta-features. Based on these meta-features the prediction of well suitable classifier is done. In this paper, in the first experiment, the prediction classifier is done by landmarking meta-features with k-NN approach. In the second experiment in addition to first experiment Win/ draw/ loss of corresponding classifiers is calculated using recommendation method and based on that the best classifier is recommended. Here the simple linear regression value of classifiers is taken into consideration. In both the experiments performance measure is the accuracy of classifier.