Dealing with Overlapping Clustering: A Constraint-based Approach to Algorithm Selection

When confronted to a clustering problem, one has to choose which algorithm to run. Building a system that automatically chooses an algorithm for a given task is the algorithm selection problem. Unlike the well-studied task of classification, clustering algorithm selection cannot rely on labels to choose which algorithm to use. However, in the context of constraint-based clustering, we argue that using constraints can help in the algorithm selection process. We introduce CBOvalue, a measure based on must-link and cannot-link constraints that quantifies the overlapping in a dataset. We demonstrate its usefulness by choosing between two clustering algorithm, EM and spectral clustering. This simple method shows an average performance increase, demonstrating the potential of using constraints in clustering algorithm selection.