Inductive Matrix Completion: No Bad Local Minima and a Fast Algorithm

Authors: Pini Zilber, Boaz Nadler

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, given entries observed uniformly at random, GNIMC recovers the underlying matrix substantially faster than several competing methods.
Researcher Affiliation Academia 1Faculty of Mathematics and Computer Science, Weizmann Institute of Science, Israel. Correspondence to: Pini Zilber <pini.zilber@weizmann.ac.il>, Boaz Nadler <boaz.nadler@weizmann.ac.il>.
Pseudocode Yes Algorithm 1 GNIMC input sampling operator PΩ, observed matrix Y , side information matrices (A, B), maximal number of iterations T, initialization (U0, V0) output rank-r (approximate) solution to PΩ( ˆX) = Y
Open Source Code Yes MATLAB and Python code implementations of GNIMC, Alt Min, GD and RGD are available at github.com/pizilber/IMC.
Open Datasets No In each simulation we construct U Rd1 r, V Rd2 r, A Rn1 d1 and B Rn2 d2 with entries i.i.d. from the standard normal distribution, and orthonormalize their columns. We then set X = AUDV B where D Rr r is diagonal with entries linearly interpolated between 1 and κ.
Dataset Splits No The paper does not specify explicit train/validation/test dataset splits. Data is synthetically generated for each simulation run.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts) used for running its experiments.
Software Dependencies Yes L-BFGS: limited-memory quasi-Newton BFGS algorithm, as implemented in MATLAB R2022a.
Experiment Setup Yes For each simulation setting, we tuned the optimal parameter out of 10 logarithmically-scaled values. The permitted values for Maxide were 10 5, ..., 10 14; for MPPF, GD and RGD were 10 2/κ, ..., 10 1 2 /κ where κ is the condition number of X ; for Scaled GD were 10 2, ..., 10 1 2 ; and for L-BFGS were 1, ..., 103. For GNIMC, in all simulations we identically set the maximal number of iterations for the inner least-squares solver to 10 if the observed error is low, PΩ(Xt) Y F Y F 10 4, and 1000 otherwise. We used the following two stopping criteria for all methods: (i) small relative observed RMSE, PΩ(Xt) Y F Y F ϵ, or (ii) small relative estimate change PΩ(Xt Xt 1) F PΩ(Xt) F ϵ. In our simulations, we set ϵ = 10 14.