Halting in Random Walk Kernels

Authors: Mahito Sugiyama, Karsten Borgwardt

NeurIPS 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We theoretically show that halting may occur in geometric random walk kernels. We also empirically quantify its impact in simulated datasets and popular graph classification benchmark datasets.
Researcher Affiliation Academia Mahito Sugiyama ISIR, Osaka University, Japan JST, PRESTO mahito@ar.sanken.osaka-u.ac.jp Karsten M. Borgwardt D-BSSE, ETH Z urich Basel, Switzerland karsten.borgwardt@bsse.ethz.ch
Pseudocode No The paper does not contain any pseudocode or algorithm blocks.
Open Source Code Yes The code and all datasets are available at: http://www.bsse.ethz.ch/mlcb/research/machine-learning/graph-kernels.html
Open Datasets Yes We collected five real-world graph classification benchmark datasets: ENZYMES, NCI1, NCI109, MUTAG, and D&D, which are popular in the graph-classification literature [13, 14]. The code and all datasets are available at: http://www.bsse.ethz.ch/mlcb/research/machine-learning/graph-kernels.html
Dataset Splits Yes The classification accuracy of each graph kernel was examined by 10-fold cross validation with multiclass C-support vector classification (libsvm2 was used), in which the parameter C for CSVC and a parameter (if one exists) of each kernel were chosen by internal 10-fold cross validation (CV) on only the training dataset.
Hardware Specification Yes We used Amazon Linux AMI release 2015.03 and ran all experiments on a single core of 2.5 GHz Intel Xeon CPU E5-2670 and 244 GB of memory.
Software Dependencies Yes All kernels were implemented in C++ with Eigen library and compiled with gcc 4.8.2. libsvm2 was used.
Experiment Setup Yes The list of parameters optimized by the internal CV is as follows: C {2 7, 2 5, . . . , 25, 27} for C-SVC, the width σ {10 2, . . . , 102} in the RBF kernel KVEH,G, the number of steps k {1, . . . , 10} in Kk , the number of iterations h {1, . . . , 10} in KWL, and λ {10 5, . . . , 10 2, λmax} in KH and KGR, where λmax = (max G,G G ) 1.