An unsupervised ensemble learning and feature ranking method in which the combiner has no information about the expert’s performance, methods, and the data with which they operate is proposed in this paper. The method uses batch processing as opposed to an online fusion method using weighted majority rule, which is the difference between this method and previous work. This method estimates the performance of experts using detection probability and false alarm probability for each of them, and then fuses the decisions of the experts. The authors describe the estimation algorithm in detail with proofs, justification of the combiner’s fusion rule, a new feature ranking algorithm, and benchmark results of the algorithm compared to the majority rule.

In experimental results, the authors first use eight different model experts to evaluate the performance of their algorithm. They also evaluate their algorithm when the Lipschitz constant is not available at the combiner. Experimental results show that the performance of the combiner improves with the number of experts and instances, and it outperforms the majority rules in all cases of reliability values.

In summary, this paper proposes a novel unsupervised ensemble method where the combiner has no context regarding the expert’s model, data, and performance; it only uses the local decisions of experts to make a final decision. This approach does not take into account any prior information about how experts use the data. Experimental results show that this unsupervised ensemble method is accurate in learning systems and extracting the importance of all features. This method is novel in its mathematical formulation and has many application areas in which it can provide significant improvement.