Boosting as a very successful classification algorithm represents a great generalization ability with appropriate ensemble diversity. It can be easily applied in the two-class classification problem. However, sequential structure prediction, in which the output is an ordered list of the labeled classes, needs to be realized by an adjusted and extended version. For that purpose the AdaBoostSeq algorithm has been introduced. It performs the multi-class classification with respect to the sequential structure of the classification target. The profile of the AdaBoostSeq algorithm is analyzed in the paper, especially its classification accuracy, using various base classifiers applied to diverse experimental datasets with comparison to other state-of-the-art methods.
Combining pattern recognition is a promising direction in designing effective classifiers. There are several approaches to collective decision-making, including quite popular voting methods where the decision is a combination of individual classifiers' outputs. The article focuses on the problem of fuser design which uses discriminants of individual classifiers to make a decision. We present taxonomy of proposed fusers and discuss some of their properties. We focus on the fuser which uses weights dependent on classifier and class number, because of a pretty low computational cost of its training. We formulate the problem of fuser learning as an optimization task and propose a solver which has its origin in neural computations. The quality of proposed learning algorithm was evaluated on the basis of several computer experiments, which were carried out on five benchmark datasets and their results confirm the quality of proposed concept.