Deep Fusion Network based Ensemble Learning Approach for Detection Intensive Classification

Abstract

Machine learning has become a popular approach for automatic detection of specific patterns. However, each learning algorithm could have its own advantages and disadvantages for dealing with special types of data, e.g. heuristic algorithms could generally lead to the production of biased classifiers, especially when learning from a small data sample that is unlikely to represent a full population. In order to address the above issue, researchers have been motivated to develop ensemble learning approaches for combining individual classifiers, towards reducing the bias of classifiers and thus advancing the classification performance. In this paper, we propose a deep fusion network based ensemble learning approach, which aims to create more complex ensembles and to achieve a strategic combination of the existing rules of fusion (e.g. majority vote, mean, median and max) in a layer-by-layer processing manner, rather than simply using a single fusion rule. The proposed ensemble learning approach is used for a special type of detection tasks, which involves one class as the target class and the other class as the default class. The performance of the proposed ensemble learning approach is evaluated using 6 UCI data sets and the results show that the proposed ensemble learning approach consistently performs very close to or better than the state-of-the art approaches of individual and ensemble learning, while the individual classifiers and the other types of ensembles varied in their performance on different data sets.