This paper presents a unifying review of a learning-based framework for kernel-based attributed graph matching. The framework, which includes as special cases the RKHS Interplator-Based Graph Matching (RIGM) and Interpolator-Based Kronecker Product Graph Matching (IBKPGM) algorithms, incorporates a general approach where no assumption is made about the adjacency structure of the graphs to be matched. Corresponding pairs of attributed adjacency matrices and attribute vectors of an input and reference graph are used as the input–output training set of a constrained multi-input multi-output multi-variable mapping to be learned. It is shown that a Reproducing Kernel Hilbert Space (RKHS) based interpolator can be used to infer this mapping. Partially constraining the inferred mapping by the generation of additional consistency input–output training pairs and the use of polynomial feature augmentation lead to improved performance. The proposed learning-based framework avoids the explicit calculation of compatibility values.
Attributed Graph Matching Using Local Descriptions
