On the effectiveness of heuristics for learning nested dichotomies: an empirical analysis

In machine learning, so-called nested dichotomies are utilized as a reduction technique, i.e., to decompose a multi-class classification problem into a set of binary problems, which are solved using a simple binary classifier as a base learner. The performance of the (multi-class) classifier thus produced strongly depends on the structure of the decomposition. In this paper, we conduct an empirical study, in which we compare existing heuristics for selecting a suitable structure in the form of a nested dichotomy. Moreover, we propose two additional heuristics as natural completions. One of them is the Best-of-K heuristic, which picks the (presumably) best among K randomly generated nested dichotomies. Surprisingly, and in spite of its simplicity, it turns out to outperform the state of the art.