Nearest neighbour group-based classification

The purpose of group-based classification (GBC) is to determine the class label for a set of test samples, utilising the prior knowledge that the samples belong to same, but unknown class. This can be seen as a simplification of the well studied, but computationally complex, non-sequential compound classification problem. In this paper, we extend three variants of the nearest neighbour algorithm to develop a number of non-parametric group-based classification techniques. The performances of the proposed techniques are then evaluated on both synthetic and real-world data sets and their performance compared with techniques that label test samples individually. The results show that, while no one algorithm clearly outperforms all others on all data sets, the proposed group-based classification techniques have the potential to outperform the individual-based techniques, especially as the (group) size of the test set increases. In addition, it is shown that algorithms that pool information from the whole test set perform better than two-stage approaches that undertake a vote based on the class labels of individual test samples.