This page is dedicated to the paper:
Fabrizio Costa, Kurt De Grave. Fast Neighborhood Subgraph Pairwise Distance Kernel. Proceedings of the 27th International Conference on Machine Learning (ICML-2010), Haifa, Israel, 2010.
Download the full text of the paper.
Software and source code
You'll need five components to run the kernel
- DMax Chemistry Assistant if you want to augment the molecular graphs. The input file is a standard MDL SDFile. After creating a session with the SDFile, you can find the knowledge base in ~/pharmaDM_user/ChemistryAssistant/sessions/session-name/data/molecules.kb
- The KB2Graphs program to convert a subset of the DMax knowledge base into GF format (use option -f). The augmentation can optionally be stripped at this point.
- The NSPDK kernel implementation itself, which takes a GF file and outputs a recursive data kernel (.rdk) file.
- An SVMDlight plugin for rdk files. (Included in the kernel download). The plugin computes the actual kernel function values.
- (Optional) SVMDlight, which is a modified, plugin-capable version of Thorsten Joachims' SVMlight.
There exists also a more convenient, stand-alone, even faster, newer version of the kernel, which takes either GSpan or SDFile formatted input. KB2Graphs can produce GSpan if you want to run the new version on augmented molecular graphs. The newer kernel version, however, does not calculate exactly the same, so the above is required to reproduce the results from the paper.