Including semi-supervision in a kernel matrix, with a view to interactive visual clustering
Pierrick Bruneau, Benoît Otjacques · HAL (Le Centre pour la Communication Scientifique Directe) · 2012
In this paper, a new kernel transformation procedure is described. It aims at incorporating a degree of supervision directly in the original pairwise similarities of a data set. The modified similarities can then be projected using a 2D kernel PCA, so as to reflect the compromise between genuine data and user knowledge, while being affordable for visualization and interaction. Such semi-supervised projections are evaluated with synthetic and real data, in the context of a simulated visual clustering task. Randomly selected subsets of elements are chosen to hold a label, thus reproducing actual user interactions. The results show the effectiveness of the method, with as few as one labelled element per class inducing tangible effects.