Conformalized matrix completion

Authors: Yu Gui, Rina Barber, Cong Ma

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical results on simulated and real data demonstrate that cmc is robust to model misspecification while matching the performance of existing model-based methods when the model is correct. In this section, we conduct numerical experiments to verify the coverage guarantee of conformalized matrix completion (cmc) using both synthetic and real datasets.
Researcher Affiliation Academia Yu Gui Rina Foygel Barber Cong Ma Department of Statistics, University of Chicago {yugui,rina,congm}@uchicago.edu
Pseudocode Yes Algorithm 1 Conformalized matrix completion (cmc)
Open Source Code Yes All the results in this section can be replicated with the code available at https: //github.com/yugjerry/conf-mc.
Open Datasets Yes We also compare conformalized approaches with model-based approaches using the Rossmann sales dataset4 (Farias et al., 2022).
Dataset Splits Yes Split the data: draw Wij i.i.d. Bern(q), and define training and calibration sets Str = {(i, j) S : Wij = 1}, and Scal = {(i, j) S : Wij = 0}.
Hardware Specification No No specific hardware details (like GPU/CPU models or memory) are mentioned in the paper for running the experiments.
Software Dependencies No The paper mentions software like 'alternating least squares (als)' and 'convex relaxation' but does not provide specific version numbers for these or any other software dependencies.
Experiment Setup Yes Throughout the experiments, we set the true rank r = 8, the desired coverage rate 1 α = 0.90, and report the average results over 100 random trials. In the synthetic setting, we generate the data matrix by M = M + E, where M is a rank r matrix, and E is a noise matrix (with distribution specified below). The low-rank component M is generated by M = κU V , where U Rd1 r is the orthonormal basis of a random d1 r matrix consisting of i.i.d. entries from a certain distribution Pu,v (specified below) and V Rd2 r is independently generated in the same manner.