Power Iterated Color Refinement

Authors: Kristian Kersting, Martin Mladenov, Roman Garnett, Martin Grohe

AAAI 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We support our theoretical results with experiments on real-world graphs with millions of edges.
Researcher Affiliation Academia Kristian Kersting TU Dortmund University {fn.ln}@cs.tu-dortmund.de Martin Mladenov TU Dortmund University {fn.ln}@cs.tu-dortmund.de Roman Garnett University of Bonn garnett@uni-bonn.de Martin Grohe RWTH Aachen grohe@informatik.rwth-aachen.de
Pseudocode Yes Algorithm 1: CGCR(A): CG for Color Refinement; Algorithm 2: COLORS(M); Algorithm 3: CHARACTMAT(M); Algorithm 4: CGCR(A): CG for Color Refinement; Algorithm 5: HCGCR(A): Hashed CGCR; Algorithm 6: PICGCR(A): Power Iterated CGCR
Open Source Code Yes we implemented naive versions of CGCR, HCGCR (denoted as Hashing and available at https://github.com/rmgarnett/fast_wl/) and PICGCR in Matlab
Open Datasets Yes Table 1: Scaling results on graphs generated from http://www.cise.ufl.edu/research/sparse/matrices/index.html (matrices were turned into graphs using A := A + AT and thresholding |A| > 0).
Dataset Splits No The paper does not specify validation splits or detailed train/validation/test partitioning for the datasets used in experiments.
Hardware Specification Yes on a single Linux machine( 4 3.4 GHz cores, 32 GB main memory)
Software Dependencies No we implemented naive versions of CGCR, HCGCR (...) and PICGCR in Matlab. Matlab version number is not specified.
Experiment Setup Yes The convergence threshold ϵ for PI was set to 10 8 and the damping factor α to 0.95. The maximum number of PI iterations was set to 500 (denoted as PIfix) resp. to min(2k, τ) (PIflex).