In-depth Analysis of Low-rank Matrix Factorisation in a Federated Setting

Authors: Constantin Philippenko, Kevin Scaman, Laurent Massoulié

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We complete our analysis with experiments on both synthetic and real data.
Researcher Affiliation Academia Inria Paris D epartement d informatique de l ENS, PSL Research University EMAIL
Pseudocode Yes Algorithm 1: Distributed Randomized Power Iteration; Algorithm 2: GD w.r.t. U with a power init.
Open Source Code Yes Code https://github.com/philipco/matrix factorization
Open Datasets Yes Real datasets. We consider three real datasets: mnist (Le Cun, Cortes, and Burges 2010), celeba-200k (Liu et al. 2015) and w8a (Chang and Lin 2011).
Dataset Splits No The paper describes how datasets are split across clients and mentions 'training dataset size' in Table 1, but it does not provide specific training/test/validation splits (percentages or counts) for model evaluation. For example, it states 'For w8a, the dataset is split randomly across clients' but no further details on train/test/validation splits for the learning task.
Hardware Specification Yes Experiments have been run on a 13th Gen Intel Core i7 processor with 14 cores.
Software Dependencies No The paper mentions 'Truncated SVD class of Scikit-learn (Pedregosa et al. 2011)' and 'svd lowrank function of Py Torch (Paszke et al. 2019)' as related work or tools used for comparison, but it does not specify version numbers for the software dependencies of its own described methodology.
Experiment Setup Yes Input: Number of iteration α in N, step-size γ. Output: (U i)N i=1. Run Algorithm 1 to compute V = (S S)αS Φ. for each client i in {1, . . . , N} without any communication do Sample a random matrix Ui 0 in Rni r. for t {1, . . . , T } do Compute UF(Ui t 1, V) = (Ui t 1V Si)V. Ui t = Ui t 1 γ UF(Ui t 1, V). ... Table 1: Settings of the experiments. ... latent dimension r 20 20 20 ... We run a single gradient descent after having sampled m = 20 random matrices Φ to take the one resulting in the best condition number κ(V). ... We run experiments w./w.o. a momentum βk = k/(k + 3), with k the iteration index.