A multiplex-network based approach for clustering ensemble selection

Performance of cluster ensemble approaches is now known to be tightly related to both quality and diversity of input base clusterings. Cluster ensemble selection (CES) refers to the process of filtering the raw set of base clusterings in order to select a subset of high quality and diverse clusterings. Most of existing CES approaches apply one index for measuring the quality and another for evaluating the diversity of clusterings. Moreover the number of clusterings to select is usually given as an input to the CES function. In this work we propose a new CES approach that allow taking into account an ensemble of quality and diversity indexes. In addition, the proposed approach computes automatically the number of clusterings to return. The basic idea is to define a multiplex network over the provided set of base clusterings. Each slice in the multiplex network is obtained by defining a proximity-graph over the set of base clusterings using a given clustering dissimilarity index. A community detection algorithm is applied to the obtained multiplex network. We then rank clusterings in each community applying an ensemble-ranking approach using different (internal) clustering quality indexes. From each community we select the base clustering ranked at the top. First experiments on benchmark datasets shows the effectiveness of the proposed CES approach.