Bucket Selection : A Model-Independent Diverse Selection Strategy for Widening

Abstract

When using a greedy algorithm for finding a model, as is the case in many data mining algorithms, there is a risk of getting caught in local extrema, i.e., suboptimal solutions. Widening is a technique for enhancing greedy algorithms by using parallel resources to broaden the search in the model space. The most important component of widening is the selector, a function that chooses the next models to refine. This selector ideally enforces diversity within the selected set of models in order to ensure that parallel workers explore sufficiently different parts of the model space and do not end up mimicking a simple beam search. Previous publications have shown that this works well for problems with a suitable distance measure for the models, but if no such measure is available, applying widening is challenging. In addition these approaches require extensive, sequential computations for diverse subset selection, making the entire process much slower than the original greedy algorithm. In this paper we propose the bucket selector, a model-independent randomized selection strategy. We find that (a) the bucket selector is a lot faster and not significantly worse when a diversity measure exists and (b) it performs better than existing selection strategies in cases without a diversity measure.