Provable Online CP/PARAFAC Decomposition of a Structured Tensor via Dictionary Learning

Authors: Sirisha Rambhatla, Xingguo Li, Jarvis Haupt

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Complementary to our theoretical results, our synthetic and real-world data evaluations showcase superior performance compared to related techniques.
Researcher Affiliation Academia 1 Computer Science Department, University of Southern California 2 Computer Science Department, Princeton University 3 Department of Electrical and Computer Engineering, University of Minnesota Twin Cities
Pseudocode Yes Algorithm 1 Tensor NOODL: Neurally plausible alternating Optimization-based Online Dictionary Learning for Tensor decompositions. Algorithm 2 Untangle Khatri-Rao Product (KRP): Recovering the Sparse factors
Open Source Code Yes Corresponding code is available at https://github.com/srambhatla/Tensor NOODL.
Open Datasets No The paper describes generating synthetic data and using specific real-world datasets (NBA Shot Pattern Dataset, Enron data) but does not provide specific links, DOIs, repositories, or formal citations for public access to these datasets as used in their experiments.
Dataset Splits No The paper describes synthetic data generation parameters and real-world data usage (weekly NBA data), but does not specify explicit train/validation/test dataset splits (percentages, sample counts, or citations to predefined splits) in the main text.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details (e.g., library or solver names with version numbers) needed to replicate the experiment.
Experiment Setup Yes Experimental set-up: We compare Tensor NOODL with online dictionary learning algorithms presented in [19] (Arora(b) (incurs bias) and Arora(u) (claim no bias)), and [20], which can be viewed as a variant of ALS (matricized) 4. We analyze the recovery performance of the algorithms across different choices of tensor dimensions J = K = {100, 300, 500} for a fixed n = 300, rank m = {50, 150, 300, 450, 600}, and the sparsity parameters α = β = {0.005, 0.01, 0.05} of factors B (t) and C (t), across 3 Monte-Carlo runs 5. Parameters ηA, ηx, τ, T, C, and R as per A.3, A.5, and A.6.