An Enhanced Univariate Discretization Based on Cluster Ensembles

Most discretization algorithms focus on the univariate case. In general, they take into account the target class or interval-wise frequency of data. In so doing, useful information regarding natural group, hidden pattern and correlation among the attributes may be inevitably lost. In response, this paper introduces a new pruning method that exploits natural groups or clusters as an explicit constraint to traditional cut-point determination techniques. This unsupervised approach makes use of cluster ensembles to reveal similarities between data belonging to adjacent intervals. To be precise, a cut-point between a pair of highly similar or related intervals will be dropped. This pruning mechanism is coupled with three different univariate discretization algorithms, with the evaluation is conducted on 10 datasets and 3 classifier models. The results suggest that the proposed method usually achieve higher classification accuracy levels, than those of the three baseline counterparts.