Riemannian Pursuit for Big Matrix Recovery

Authors: Mingkui Tan, Ivor W. Tsang, Li Wang, Bart Vandereycken, Sinno Jialin Pan

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

Reproducibility Variable Result LLM Response
Research Type Experimental experimental results show that it substantially outperforms existing methods when applied to large-scale and ill-conditioned matrices.
Researcher Affiliation Academia 1 School of Computer Science, The University of Adelaide, Ingkarni Wardli North Terrace Campus 5005, Australia 2 Center for Quantum Computation & Intelligent Systems, University of Technology Sydney, Australia 3 Department of Mathematics, University of California, San Diego, USA 4 Department of Mathematics, Princeton University, Fine Hall, Washington Road, Princeton NJ 08544-1000, USA 5 Institute for Infocomm Research, 1 Fusionopolis Way, #21-01 Connexis (South) 138632, Singapore
Pseudocode Yes Algorithm 1 RP: Riemannian Pursuit for MR. and Algorithm 2 NRCG(Xintial,r, ϵin) for solving (2).
Open Source Code No The paper does not provide an explicit statement or link to the source code for the proposed Riemannian Pursuit (RP) method.
Open Datasets Yes Movie Lens with 10M ratings (denoted by Movie-10M) (Herlocker et al., 1999) and Netflix Prize dataset (KDDCup, 2007).
Dataset Splits Yes we report the testing RMSE of different methods over 10 random 80/20 train/test partitions as explained in (Laue, 2012).
Hardware Specification Yes All the experiments (except for GECO) are conducted in Matlab on a PC installed a 64-bit operating system with an Intel(R) Core(TM) i7 CPU (2.80GHz with single-thread mode) and 24GB memory.
Software Dependencies No The paper states that experiments were conducted 'in Matlab' but does not provide a specific version number for Matlab or any other software dependencies with version numbers.
Experiment Setup Yes We set λF = 0.01, ϵout = 10 5 for the stopping conditions and η = 0.65 for RP. For APG, we set the trade-off parameter λ = 10 3σmax, where σmax is the largest singular value of A (b). For all the fixed-rank methods, we set r = br = 50.