Efficient Sparse Low-Rank Tensor Completion Using the Frank-Wolfe Algorithm

Authors: Xiawei Guo, Quanming Yao, James Kwok

AAAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that the proposed algorithm is more accurate, much faster and more scalable than the state-of-the-art.
Researcher Affiliation Academia Xiawei Guo, Quanming Yao, James T. Kwok Department of Computer Science and Engineering Hong Kong University of Science and Technology Clear Water Bay, Hong Kong {xguoae, qyaoaa, jamesk}@cse.ust.hk
Pseudocode Yes Algorithm 1 Frank-Wolfe (FW) algorithm. Algorithm 2 Reducing the size of {Ud, Σd, Vd}d [D]. Algorithm 3 Fast FW algorithm for tensor completion (FFWTensor).
Open Source Code No The paper provides links to the code for *other* methods (e.g., Geom CG, Rprecon, ADMM, Fal RTC, TMac, TTN), but does not provide a link or statement about the availability of the source code for their own proposed FFWTensor algorithm.
Open Datasets Yes The first one is Climate Net8, which is constructed from the 5 5 latitude-longitude gridded climate data set. ... 8http://www.nd.edu/dial/software/climate Net.zip. The second one is You Tube data set9 (Tang, Wang, and Liu 2009), with 15, 088 users and 5 types of Boolean interactions. ... 9http://leitang.net/data/youtube-data.tar.gz
Dataset Splits Yes 5% of the 2901 3000 3 data tensor entries are randomly sampled as the training set, another 20% as validation set (for parameter tuning), and the rest as testing set. (Color Image Inpainting) 5% of the entries in the 1773 1773 7 data tensor are sampled as observed, another 20% for validation and the rest for testing. (Climate Net) From the 15088 15088 5 data tensor, we randomly sample 0.8% of the entries as observed, another 0.1% for validation and 0.15% for testing. (You Tube full set) From the 1509 1509 5 data tensor, we sample 5% of the entries as observed, another 20% for validation, and the rest for testing. (You Tube subset)
Hardware Specification Yes Experiments are performed on a PC with Intel i7 CPU and 16GB RAM.
Software Dependencies No The paper states 'All the codes are in Matlab.' but does not specify particular Matlab versions or any other software libraries with version numbers.
Experiment Setup Yes We use the same stopping criterion that allows each to run for 1000 seconds. (General Experiments) which is triggered when the size exceeds K = 100. (Basis Reduction) 0.6% of the entries in A are randomly sampled as observed, 0.1% for validation, and another 0.1% for testing. (Synthetic Data)