10 June 2020
New optimised ensemble classifier for classification and prediction
A new ensemble classifier method has been released via GitHub and it is available to researchers and general community to test and use it for prediction and classification.
The main features of this ensemble classifier are:
automatically produces best ensemble on given data
uses optimal clustered input data space
uses input data space clustered between 1 cluster and N (total samples) - all input space
selects and uses best multiple types of base classifiers
makes final decision from optimised diverse base classifiers
guaranteed highest accuracy compared to any types of single classifier
The new ensemble classifier has been tested on benchmark datasets collected from the UCI machine learning repository. The average classification accuracy on benchmark datasets was 8.58%, 5.96%, 6.80% and 7.39% higher than the accuracy obtained by classifiers such as neural networks, support vector machines, k-nearest neighbor and decision trees respectively.
The accuracy was also compared with a recently published and top performing ensemble classifiers including ensemble that uses a random subspace through clustering and incorporates a rule-based accuracy and diversity. The new ensemble classifier outperformed most existing ensemble classifiers.
Lead investigator Prof Brijesh Verma said the main benefit of new ensemble classifier is that it can automatically find the best optimised ensemble which can produce highest accuracy on given dataset.
"We don’t need to go through a trial and error based time consuming method to find the best classifier or most suitable classifier," he said.
The full code is available at GitHub: https://github.com/bvermaqld/ARCDP
This project was funded by ARC discovery project grant (Project ID: DP160102639).
Media contact:
Brijesh Verma, b.verma@cqu.edu.au
0732951156