Meta-Learning and Bi-Tier Selection Based Classification Algorithm Recommendation

Abstract

Aiming at the problem of how to select the best algorithm for a given classification problem, a meta-learning and Bi-Tier selection based classifier recommendation (MLBCR) is proposed. Firstly, a set of meta-features is extracted to characterize the classification dataset by combining meta-learning. Secondly, a comprehensive evaluation index is proposed to comprehensively evaluate the performance of 11 candidate classification algorithms on each classification dataset, and the algorithm with the best evaluation is defined as the best algorithm for the dataset. Then, the meta-feature vector corresponding to the training dataset and the best algorithm are stored in the MySQL database in the form of tuples. Finally, the dynamic interval is set according to the similarity between the features, the similar data sets of the new data set are identified, the algorithms corresponding to the similar data sets are obtained from the database, and the algorithm with the best performance is selected and recommended to the new classification dataset. The experimental results show that the recommendation accuracy of the proposed algorithm recommendation model and the classification accuracy of the recommended best algorithm are improved by 2.42% and 1.18% respectively, which provides a more applicable solution for solving practical classification problems