Uniform Hypergraph Partitioning: Provable Tensor Methods and Sampling Techniques

Authors: Debarghya Ghoshdastidar, Ambedkar Dukkipati

JMLR 2017 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We also complement our theoretical findings by elaborate empirical comparison of various hypergraph partitioning schemes. Section 6 presents a wide variety of empirical studies that (i) validate our theoretical findings regarding relative merits of TTM over previously studied algorithms, (ii) compare spectral methods to other hypergraph partitioning algorithms, including popular hMETIS tool (Karypis and Kumar, 2000), and (iii) weigh the merits of hypergraph partitioning in subspace clustering applications, including benchmark problems.
Researcher Affiliation Academia Debarghya Ghoshdastidar EMAIL Ambedkar Dukkipati EMAIL Department of Computer Science & Automation Indian Institute of Science Bangalore 560012, India
Pseudocode Yes Algorithm TTM : Spectral relaxation of tensor trace maximisation problem; Algorithm Sampled TTM : TTM where a sampled set of edge weights are observed; Algorithm Tetris : TTM with iterative sampling for subspace clustering
Open Source Code Yes Codes are available at: http://sml.csa.iisc.ernet.in/SML/code/Feb16Tensor Trace Max.zip and also on the personal webpage of first author.
Open Datasets Yes The Hopkins 155 database (Tron and Vidal, 2007) contains a number of videos capturing motion of multiple objects or rigid bodies.
Dataset Splits No No explicit training/test/validation splits are provided. For synthetic data, the paper describes generating 'n/k points' from 'k=5 subspaces'. For Hopkins 155, it mentions '120 sequences, each containing two motions, and 35 three motion sequences' but no explicit splits are detailed.
Hardware Specification Yes We note that the reported time is based on the fact that we have used Matlab implementations of the algorithms, run on a Mac OS X operating system with 2.2 GHz Intel Core i7 processor and 16 GB memory.
Software Dependencies No The paper mentions 'Matlab implementations' and 'Mac OS X operating system' but does not provide specific version numbers for Matlab or any other libraries/solvers.
Experiment Setup Yes For existing methods, we fix the parameters as mentioned in Park et al. (2014). For Tetris and SGC, the parameters are set to the same values as SCC, where c = 100k and σ as in (31) is determined by the algorithm. In case of uniformly sampled TTM, we fix σ to be same as the value determined by Tetris.