Active graph matching based on pairwise probabilities between nodes

Abstract

We propose a method to perform active graph matching in which the active learner queries one of the nodes of the first graph and the oracle feedback is the corresponding node of the other graph. The method uses any graph matching algorithm that iteratively updates a probability matrix between nodes (Graduated Assignment, Expectation Maximisation or Probabilistic Relaxation). The oracle's feedback is used to update the costs between nodes and arcs of both graphs. We present and validate four different active strategies based on the probability matrix between nodes. It is not needed to modify the code of the graph-matching algorithms, since our method simply needs to read the probability matrix and to update the costs between nodes and arcs. Practical validation shows that with few oracle's feedbacks, the algorithm finds the labelling that the user considers optimal because imposing few labellings the other ones are corrected automatically. © 2012 Springer-Verlag Berlin Heidelberg.