CBMI2015 paper on efficient pairwise constraints propagation for graph clustering

Date de publication : Jun 16, 2015 8:28:19 AM

This paper "Semi-supervised Spectral Clustering with automatic propagation of pairwise constraints" proposed at CBMI2015 presents how to propagate similar/not similar constraints between objects of a dataset. The aim is to help graph clustering methods such as Spectral Clustering, to better identify clusters at an early stage of the annotation process and limiting time consuming human manual data annotation. This paper encourages to use two state of the art propagation methods and proposes a new one that significantly amplify propagation thanks to a cascade effect with the two others.

As a conclusion, the proposed framework allows to efficiently annotate a dataset with the following properties :

_annotation relies on very simple flags that compare object pairs (similar /not similar). This is simpler than absolute class annotation.

_the oracle (human annotator) is less and more efficiently involved, only annotating pairs that could not be annotated automatically.

_computing time is reduced since many clustering steps are avoided. .