Extending Tree Kernels with Topological Information

The definition of appropriate kernel functions is crucial for the performance of a kernel method. In many of the state-of-the-art kernels for trees, matching substructures are considered independently from their position within the trees. However, when a match happens in similar positions, more strength could reasonably be given to it. Here, we give a systematic way to enrich a large class of tree kernels with this kind of information without affecting, in almost all cases, the worst case computational complexity. Experimental results show the effectiveness of the proposed approach.

• Type: Journal
• Year: 2011
• Where: ICANN
• Authors: Fabio Aiolli, Giovanni Da San Martino, Alessandro Sperduti
• Keywords: Computational Complexity, ICANN 2011, Kernel Functions, Kernel Method, Neural Networks