Rank-One Matrix Pursuit for Matrix Completion

Authors: Zheng Wang, Ming-Jun Lai, Zhaosong Lu, Wei Fan, Hasan Davulcu, Jieping Ye

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

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically evaluate the proposed algorithm on many real-world large-scale datasets. Results show that our algorithm is much more efficient than state-of-the-art matrix completion algorithms while achieving similar or better prediction performance.
Researcher Affiliation Collaboration The Biodesign Institue, Arizona State University, Tempe, AZ 85287, USA Department of Mathematics, University of Georgia, Athens, GA 30602, USA Department of Mathematics, Simon Fraser University, Burnaby, BC, V5A 156, Canada Huawei Noah s Ark Lab, Hong Kong Science Park, Shatin, Hong Kong School of Computing, Informatics, and Decision Systems Engineering, Arizona State University, Tempe, AZ 85281, USA
Pseudocode Yes Algorithm 1 Rank-One Matrix Pursuit (R1MP) and Algorithm 2 Economic Rank-One Matrix Pursuit (ER1MP)
Open Source Code No The paper provides links to third-party code used for comparison, but does not provide a specific link or explicit statement for its own methodology's source code.
Open Datasets Yes The Jester datasets were collected from a joke recommendation system... The Movie-Lens datasets were collected from the Movie Lens website4. (Footnote 4: http://movielens.umn.edu)... We use the following benchmark test images: Lenna, Barbara, Clown, Crowd, Girl, Man3. (Footnote 3: Images are downloaded from http://www.utdallas. edu/ cxc123730/mh_bcs_spl.html)
Dataset Splits Yes In the following experiments, we randomly split the ratings into training and test sets. Each set contains 50% of the ratings. We compare the prediction results from different methods... In each run, we randomly exclude 50% of the pixels in the image, and the remaining ones are used as the observations.
Hardware Specification Yes The experiments are run in a PC with WIN7 system, Intel 4 core 3.4 GHz CPU and 8G RAM.
Software Dependencies No The paper mentions software like MATLAB, C++, and PROPACK (an SVD package) but does not provide specific version numbers for any of these dependencies.
Experiment Setup Yes As the image matrix is not guaranteed to be low rank, we use the rank 200 for the estimation matrix for each experiment. The JS algorithm does not explicitly control the rank, thus we fix its number of iterations to 2000... for other algorithms we use the same rank for the estimated matrices; the values of the rank are {10, 10, 5, 10, 10, 20} for the six corresponding datasets.