Supervised Learning with Unsupervised Output Separation
Nathalie Japkowicz · 2002
In supervised learning approaches, the output labels are imposed by the knowledge engineer who prepared the data. While knowing the labels of a data set is quite useful, in cases where data points belonging to very different data distributions are agglomerated in the same class, a learning algorithm can have difficulties modeling these classes accurately. In such cases, it should be useful to separate the main classes into a number of more homogeneous subclasses. This paper assumes that the above problem is quite common and describes a simple combination method that attempts to fix it. It then tests the approach on 5 domains taken from the UCI Repository. The results show that in three out of five cases, the approach has a positive effect, in one case, it breaks even and in the fifth case, it degrades the previously established performance.