Can subclasses help a multiclass learning problem?

There are many issues that are extensively studied in a multiclass learning problem, e.g. classifier selection, data balancing, training schemes, etc. In this paper, we introduce the subclass partition and present it as a novel factor that will influence the multiclass learning performance. In many multiclass learning problems, the implicit subclass subsumption information is often ignored. This paper studies the role of the subclass and outlines the connection between the discrimination boundary and the subclass partition of the classifier. The paper investigates the use of the subclass partition to help design the better classifier architecture and formalizes the subclass partition by introducing the partition space and the partition search tree. It also proposes an algorithm to heuristically search for the optimal subclass partition. We apply the proposed scheme to a 3-class occupant classification problem, which includes 16 subclasses. The experiments are positive to demonstrate that we can improve the multiclass learning performance by searching for the optimal output class partitions.