A comparative study on clustering-based classification algorithms

For most classifiers, they are based on underlying assumptions or models. If the model or assumption matches the true data distribution, the accuracy is high, and vice versa. That is the problem of model-misfit. This paper uses clustering or cluster analysis in classification to solve the problem. For this reason, we propose a general clustering-based classification framework. In the framework, clustering algorithm is used to re-collect the data, like a filter. Thus, the diversities between classes are weakened. By that, Model-misfit is no problem any more. In our framework, almost all clustering and classification algorithms can be integrated together for the better performance. In this paper, we present an empirical study on four clustering-based classification methods. On complex data (e.g., non-linear), experimental results show that the clustering-based approach can improve the performance of the traditional classifier, especially for simple classifiers (e.g., k-NN).