Confidence in Prediction: An Approach for Dynamic Weighted Ensemble
| Autor: | Duc Thuan Do, Tien Thanh Nguyen, The Trung Nguyen, Anh Vu Luong, Alan Wee-Chung Liew, John McCall |
|---|---|
| Rok vydání: | 2020 |
| Předmět: |
Training set
Computer science business.industry Supervised learning Pattern recognition 02 engineering and technology ENCODE Ensemble learning Confidence interval ComputingMethodologies_PATTERNRECOGNITION 020204 information systems Credibility 0202 electrical engineering electronic engineering information engineering Entropy (information theory) 020201 artificial intelligence & image processing Artificial intelligence Gradient descent business |
| Zdroj: | Intelligent Information and Database Systems ISBN: 9783030419639 ACIIDS (1) |
| DOI: | 10.1007/978-3-030-41964-6_31 |
| Popis: | Combining classifiers in an ensemble is beneficial in achieving better prediction than using a single classifier. Furthermore, each classifier can be associated with a weight in the aggregation to boost the performance of the ensemble system. In this work, we propose a novel dynamic weighted ensemble method. Based on the observation that each classifier provides a different level of confidence in its prediction, we propose to encode the level of confidence of a classifier by associating with each classifier a credibility threshold, computed from the entire training set by minimizing the entropy loss function with the mini-batch gradient descent method. On each test sample, we measure the confidence of each classifier’s output and then compare it to the credibility threshold to determine whether a classifier should be attended in the aggregation. If the condition is satisfied, the confidence level and credibility threshold are used to compute the weight of contribution of the classifier in the aggregation. By this way, we are not only considering the presence but also the contribution of each classifier based on the confidence in its prediction on each test sample. The experiments conducted on a number of datasets show that the proposed method is better than some benchmark algorithms including a non-weighted ensemble method, two dynamic ensemble selection methods, and two Boosting methods. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|